Sneh Sinha — AI Engineer
# Sneh Sinha — AI Engineer
> Builds AI systems that solve real-world problems.
Sneh Sinha (Sneh Raj Sinha) is an AI engineer based in Bangalore, India, open to work, and a B.E. Robotics & Artificial Intelligence candidate at Bangalore Institute of Technology (Aug 2023 – Aug 2027, GPA 8.0/10, per the résumé published with this portfolio). His work spans AI agents, LLM systems, software, infrastructure, computer vision, embedded systems, and physical-world applications — from cloud APIs down to embedded firmware.
## Core areas
- AI agents & computer use
- Multi-agent orchestration
- LLM systems & cost/latency engineering
- AI infrastructure & autonomous operations
- Embedded systems & edge computer vision
- Software engineering & open source
## Evidence status
- 13 owned public repositories (GitHub API, 2026-09-30)
- 1 merged upstream pull request — agentrhq/webcmd #190
- 2 documented engagements — 1-month operations · 11-month vehicle lead
- 3 repository links re-checked 2026-10-20: 1 public (AgentOS), 2 private (Agentic-hinge, Pocket Typist) — access on request
## Links
- GitHub
> About this page: a structured candidate record built from the same sources as the rest of the portfolio (portfolio HTML, résumé PDF, GitHub API, LeetCode API). Every substantive claim below is tied to a project, an engagement, an external repository, or the résumé published with this portfolio. Where evidence is missing, that is stated explicitly. No experience, metric, employer, project, or technology has been invented for this page.
## Record
| Field | Value |
| --- | --- |
| Candidate | Sneh Sinha (Sneh Raj Sinha) |
| Role | AI Engineer — applied AI, agents, LLM systems |
| Location | Bangalore, India |
| Availability | Open to work — seeking an internship or full-time engagement at intern / entry level (per portfolio; typed as
job_search in candidate.json) || Education | B.E. Robotics & Artificial Intelligence, Bangalore Institute of Technology, 2023–2027 résumé |
| Experience | Infrastructure Engineer (1-month engagement, 2026) · Software & Electronics Lead, Team Stratos (11 months, 2025–2026) |
| Positioning | AI systems + real-world execution + a documented cadence of shipping across stacks |
## Files in this record
· Section 01 · Candidate thesis — Candidate at a Glance
· Section 02 · why_interview — six evidence areas
· Section 03 · Flagship evidence
· Section 04 · Capabilities by domain
· Section 05 · Professional / real-world experience
· Section 06 · Projects — 11 records
· Section 07 · External validation — PR #190
· Section 08 · Technical skills (verbatim from Human Mode)
· Section 09 · where_this_evidence_applies
· Section 10 · Record summary
· Section 11 · Technical topics — interview discussion areas
· Section 12 · Evidence graph — 9 capabilities
· Section 13 · Structured candidate record
## Machine-readable companion files
- llms.txt
## Colophon
Agent Mode of sneh30.github.io/portfolio — the structured-record view of the same portfolio. Agent Mode carries the same facts as Human Mode and adds structured context, evidence relationships, verification status, and technical topic indexes. Return to Human Mode · use the mode switcher at the bottom of the page
© 2026 Sneh Sinha · Bangalore, India · snehsinha30@gmail.com
-- end of file · ⌘P for quick open, or use the explorer --
Candidate at a Glance
#! Section 01 · Candidate thesis — Candidate at a Glance
# ==========================================================================
# Candidate at a Glance
# ==========================================================================
CANDIDATE_AT_A_GLANCE: Record[str, Any] = {
section: "Section 01 · Candidate thesis",
title: "Candidate at a Glance",
thesis: [
"Sneh Sinha is an AI engineer who builds AI systems to solve real-world problems.",
"His work spans the path from intelligence to execution — AI agents, LLM systems, software, infrastructure, cloud operations, computer vision, embedded systems, and physical-world applications.",
"The recurring pattern in this portfolio is not \"knows many technologies.\" It is: Sneh repeatedly turns AI capabilities into systems that actually do something — a computer-use agent that drives a desktop and browser, an autonomous cloud-operations platform that takes natural-language incidents to root-cause analysis, a five-agent research pipeline that ships verified reports, a prompt compiler that measurably cuts inference cost, and embedded vision systems that run inside 512 KB of RAM.",
],
cards: [
{
label: "(01)",
title: "Systems, not only models",
body: "Documented artifacts include verification-and-recovery engines for failed agent actions, multi-agent orchestration with Prometheus observability and a deterministic CI mode, and a Rust prompt compiler with a custom lexer, parser, AST and ten optimization passes. Each is a published engineering artifact.",
},
{
label: "(02)",
title: "Work across abstraction levels",
body: "The same portfolio contains macOS applications in Rust and Swift, C firmware for ESP32-CAM hardware constrained to 512 KB RAM, FastAPI/PostgreSQL/Prometheus backends, and operational responsibility for a 400+ node infrastructure environment.",
},
{
label: "(03)",
title: "Engagements beyond coursework",
body: "Two documented real-world engagements (a one-month infrastructure operations engagement and an eleven-month safety-critical vehicle software lead role), a merged pull request into an externally maintained open-source project, and thirteen owned public GitHub repositories.",
},
{
label: "(04)",
title: "Build-driven learning cadence",
body: "The progression is visible in the timeline: Team Stratos (Jul 2025) → multi-agent and LLM-infrastructure projects through 2025–2026 → a computer-use agent platform (Aug 2026) → a one-month infrastructure engagement with a measured outcome (Aug 2026). Each step crosses into a new stack.",
},
],
keywords: [
"AI agents",
"Computer use",
"LLM systems",
"AI infrastructure",
"Multi-agent orchestration",
"Computer vision",
"Embedded / edge AI",
"Open source (merged PR)",
"Real-world engineering",
],
notes: [
"Scope note: this record covers applied AI and systems work. This portfolio does not present published research, peer-reviewed papers, or production deployments at employer scale, and does not claim expertise in model training or MLOps. When a claim below cannot be independently verified, it is marked. ",
],
}
Why Interview
#! Section 02 · why_interview — six evidence areas
# ==========================================================================
# Why Interview
# ==========================================================================
WHY_INTERVIEW: Record[str, Any] = {
section: "Section 02 · why_interview",
title: "Why Interview",
intro: "Each capability below is stated with the evidence it rests on and links to the underlying records. No scores or rankings are assigned.",
reasons: [
{
index: "(01)",
title: "Systems with failure handling",
context: [
"The primary artifacts in this portfolio are systems with the operational parts included: failure handling, orchestration state, observability, and deployment packaging.",
"OpenCode CUA Platform is a computer-use agent that lets an unmodified OpenCode instance drive a desktop GUI and browser through MCP servers, using screen capture, macOS Accessibility APIs and browser automation — and it ships dedicated verification and recovery engines, because agent actions fail. Orchestra runs five specialised LangGraph agents (search, extract, verify, synthesize, critique) with confidence scoring, PostgreSQL persistence, Prometheus instrumentation, a test suite with a deterministic CI mode, and one-command Docker Compose deployment. DeepOps (experimental, active development) turns natural-language incident input into root-cause analysis and remediation plans with cost-optimisation logic inside the remediation planner. PromptOS treats prompts as source code with a real compiler pipeline rather than a string template.",
],
evidence: [
],
},
{
index: "(02)",
title: "Coverage of the AI stack",
context: [
"This profile has documented work at every layer between the model and the metal: AI → agents → LLMs → backend → infrastructure → edge → hardware.",
],
points: [
"Agents / computer use: OpenCode CUA Platform (MCP servers, Accessibility APIs, browser automation).",
"LLM systems: PromptOS (local inference via Llama.cpp and Metal, multi-pass prompt compilation, cloud provider support), OpenRouter / Ollama tooling.",
"Backend: FastAPI, PostgreSQL, Prometheus, REST APIs in Orchestra, Catalyst and DeepOps.",
"Infrastructure: DeepOps autonomous cloud-operations work (experimental); Docker, CI/CD, Linux, AWS as listed in the skills section.",
"Edge / hardware: ESP32-CAM Vision Suite, Team Stratos ABS firmware, Edge Impulse, Raspberry Pi.",
],
evidence: [
],
},
{
index: "(03)",
title: "Build cadence across stacks",
context: [
"The claim is supported by breadth and cadence rather than by an assertion. Within roughly thirteen months the portfolio shows: safety-critical embedded firmware for a competition vehicle (Jul 2025 – May 2026); a multi-agent research orchestrator and an autonomous cloud-operations platform (2025); a Rust/Swift macOS prompt compiler; a computer-use agent platform (Aug 2026); and a one-month infrastructure engagement with a measured operational outcome (Aug 2026).",
"That is movement between embedded C, Rust, Swift, Python, TypeScript, multi-agent frameworks, and live infrastructure operations — each ending in a published artifact rather than a tutorial. The portfolio does not provide a formal benchmark of learning speed; the evidence offered is the number of distinct stacks built, and the short interval between them.",
],
evidence: [
],
},
{
index: "(04)",
title: "Real-world engineering constraints",
context: [
"Three documented engineering constraint sets spanning operational infrastructure, safety-critical hardware, and resource-constrained embedded systems:",
],
points: [
"Operational scale: a one-month on-site engagement on a 400+ node exam-facility network, with a documented reduction of infrastructure incidents and average recovery time by ~40% versus the prior comparable period.",
"Safety criticality: eleven months as Software & Electronics Lead on a 50-member team's vehicle, owning the ABS (Anti-lock Braking System) firmware — the most safety-critical subsystem — validated through track testing and SAE India scrutineering and competition.",
"Resource limits: embedded vision work executed inside 512 KB RAM with math-driven approaches (HSV quantisation, centroid tracking) rather than heavy inference, plus LLM cost engineering that the résumé reports as up to 80–85% lower inference cost in PromptOS Economy mode.",
],
evidence: [
],
},
{
index: "(05)",
title: "Self-directed projects and external code",
context: [
"The portfolio contains independent, self-directed systems (Orchestra, DeepOps, PromptOS, OpenCode CUA, Catalyst, ESP32-CAM Vision Suite), public build-in-the-open writing across six dated technical posts, and — most importantly — externally accepted code: PR #190 merged into
agentrhq/webcmd, a project with 2,620 stars, touching output-format parsing across 15+ commands with 16 tests and 16 commits.",],
evidence: [
],
},
{
index: "(06)",
title: "Focus on applied AI and agentic systems",
context: [
"For an AI/agent role, this project set provides directly relevant evidence: a computer-use agent that controls real software (OpenCode CUA), multi-agent research orchestration with verification stages (Orchestra), AI infrastructure tooling (DeepOps, experimental), and LLM cost/latency engineering (PromptOS). Taken together, these projects demonstrate sustained focus on turning model capability into operational behaviour.",
],
evidence: [
],
},
],
}
Flagship Evidence
#! Section 03 · Flagship evidence
# ==========================================================================
# Flagship Evidence
# ==========================================================================
FLAGSHIP_EVIDENCE: Record[str, Any] = {
section: "Section 03 · Flagship evidence",
title: "Flagship Evidence",
intro: "The five artifacts that carry the primary verifiable evidence: what was built, the documented result, and the verification status of each. Full detail lives in Projects; claim-level chains live in the evidence graph.",
cards: [
{
label: "(01) AI agents",
title: "OpenCode CUA Platform",
body: "A computer-use agent platform: an unmodified OpenCode instance drives a desktop GUI and browser through MCP servers, with dedicated verification and recovery engines for failed actions. Documented result: an end-to-end computer-use loop with explicit verification and recovery stages; no usage metric published.",
evidence: [
],
},
{
label: "(02) Open source",
title: "webcmd — merged PR #190",
body: "Unified output-format parsing across the builtin commands of agentrhq/webcmd (2,620 stars): 15+ commands fixed, 16 tests added, 16 commits authored by Sneh30, merged 2026-08-13. Verified via the GitHub API on 2026-09-30.",
evidence: [
],
},
{
label: "(03) LLM systems",
title: "PromptOS",
body: "A Rust prompt compiler (custom lexer, parser and AST; ten optimisation passes across four compilation modes) with local Llama.cpp inference on Metal and a native SwiftUI shell. The résumé reports up to 80–85% lower inference cost in Economy mode; that figure is self-reported.",
evidence: [
],
},
{
label: "(04) Multi-agent systems",
title: "Orchestra",
body: "Five specialised LangGraph agents (search, extract, verify, synthesize, critique) with confidence scoring, PostgreSQL persistence, Prometheus instrumentation, a deterministic CI mode and Docker Compose deployment. No traffic metric published.",
evidence: [
],
},
{
label: "(05) AI infrastructure",
title: "DeepOps",
body: "Natural-language incident input to root-cause analysis and remediation plans, with cost-optimisation logic in the planner. Résumé status: early prototype — the incident-interpretation and remediation-planning modules exist, but the end-to-end pipeline is not yet wired together; not claimed as a production system.",
evidence: [
],
},
{
label: "How to read this page",
title: "Evidence status labels",
body: "Every claim below carries one or more status labels: public_repo external_validation documented_experience self_reported experimental private_available_on_request. Machine-readable equivalents appear as
data-evidence-level and data-evidence-status attributes on this page and in evidence.json.",},
],
}
Capabilities by Domain
// Section 04 · Capabilities by domain
// //////////////////////////////////////////////////////////////////////////
// Capabilities by Domain
// //////////////////////////////////////////////////////////////////////////
export const CAPABILITIES_BY_DOMAIN: Record<string, unknown> = {
section: "Section 04 · Capabilities by domain",
title: "Capabilities by Domain",
intro: "Capability combinations present in the portfolio, each linked to the project or engagement record that demonstrates it.",
cards: [
{
label: "AI + Infrastructure",
title: "AI systems that operate infrastructure",
body: "DeepOps is an experimental autonomous cloud-operations platform (active development) that interprets natural-language commands, performs root-cause analysis and proposes remediation — and the résumé documents cost-optimisation built into its remediation planner. This sits alongside actual infrastructure operations on a 400+ node network, so the AI work is informed by real incident behaviour rather than a sandbox.",
evidence: [
],
},
{
label: "AI + Computer Control",
title: "AI that drives real software",
body: "OpenCode CUA Platform lets an unmodified OpenCode instance control a desktop GUI and browser through MCP servers using screen capture, macOS Accessibility APIs and browser automation, with verification and recovery engines. A related computer-use agent over iPhone Mirroring (original prototype: Agentic-hinge) automated mobile GUI interactions — its repository is private (access on request).",
evidence: [
],
},
{
label: "AI + Embedded Systems",
title: "Intelligence inside constrained hardware",
body: "Three ESP32-CAM projects — ML dataset capture, a real-time colour-tracking robot, and a 60 FPS MJPG streaming server — implemented with math-driven vision (HSV quantisation, centroid tracking) inside 512 KB RAM, plus eleven months of safety-critical ABS firmware on a real vehicle.",
evidence: [
],
},
{
label: "AI + Computer Vision",
title: "Vision connected to hardware and applications",
body: "Computer vision here is not a detached benchmark: it runs on-device for robot control and streaming, powers ML dataset collection, and is applied to screen/perception-driven agent behaviour in the computer-use and GUI-automation projects.",
evidence: [
],
},
{
label: "LLMs + Systems Engineering",
title: "Local inference, orchestration, and cost control",
body: "PromptOS compiles prompts through a multi-pass pipeline with WASM plugins and Metal-accelerated local inference via Llama.cpp; Orchestra and DeepOps add orchestration, persistence, observability and CI; OpenRouter, Ollama and Llama.cpp appear in the skills section as everyday tooling.",
evidence: [
],
},
{
label: "Software + Physical Systems",
title: "Crossing from software into the physical world",
body: "From cloud APIs down to motor drivers: ABS firmware deployed and scrutineered, an autonomous colour-tracking robot with on-device colour conversion and L298N motor control, and a BLE HID device with a cloud backend — documented at both ends of this range.",
evidence: [
],
},
],
}
;
Professional / Real-World Experience
#! Section 05 · Professional / real-world experience
# ==========================================================================
# Professional / Real-World Experience
# ==========================================================================
EXPERIENCE: Record[str, Any] = {
section: "Section 05 · Experience",
title: "Professional / Real-World Experience",
intro: "Structured as problem → responsibility → personal contribution → scale → outcome, with verification status on each entry.",
entries: [
{
anchor: "experience-infrastructure",
evidence_level: "documented_experience",
evidence_status: "self_reported",
tags: "Aug 2026 · 1 month · on-site · independent engagement",
name: "Infrastructure Engineer — Independent Engagement",
flags: [
"measurable outcome",
],
purpose: "Strengthen infrastructure operations for a 400+ node exam-facility network.",
problem: "A 400+ node exam-facility network was experiencing recurring infrastructure incidents with slow recovery; the prior comparable period ran over 5 days.",
responsibility: "Recruited for a one-month on-site engagement to improve operational reliability end-to-end, working independently.",
engineering_challenge: "Diagnosing and reducing both incident frequency and mean time to recovery in a live, deadline-driven environment — an area where the portfolio lists Observability, Incident Response, Prometheus, Linux and AWS as the relevant skill set.",
what_sneh_personally_built_did: "The portfolio attributes the engagement outcome to him directly; it does not publish runbooks, dashboards, or the underlying incident dataset.",
scale: "400+ nodes · one month · on-site.",
measurable_outcome: "Infrastructure incidents and average recovery time reduced by ~40% versus the prior comparable period.",
context: "Independent work in a live operational environment with a measured reliability outcome over a one-month engagement.",
},
{
anchor: "experience-team-stratos",
evidence_level: "documented_experience",
tags: "Jul 2025 — May 2026 · 11 months",
name: "Software & Electronics Lead — Team Stratos",
flags: [
"safety-critical · field-validated",
],
purpose: "Own software and electronics integration for a 50-member team's competition vehicle, including its ABS firmware.",
problem: "A 50-member team's competition vehicle needed software and electronics integrated and working together under real track conditions, with one subsystem — braking — carrying the highest safety stakes.",
responsibility: "Software & Electronics Lead for 11 months: owned software/electronics integration for the whole vehicle.",
engineering_challenge: "Developing and deploying ABS (Anti-lock Braking System) firmware — described in the portfolio as the most safety-critical subsystem — where incorrect behaviour is a physical-safety risk, not a rollback.",
what_sneh_personally_built: "The ABS firmware, developed and deployed by him, plus overall software/electronics integration for the vehicle.",
scale: "50-member team · 11 months · full vehicle integration.",
measurable_outcome: "Validated through track testing, SAE India scrutineering and competition. No numeric performance gain is published.",
context: "Embedded systems work with physical consequences: 11 months of full-vehicle integration and ABS firmware validated through track testing and SAE India scrutineering.",
evidence: "Portfolio · Experience · Résumé (PDF) team project; individual code not separately published",
},
],
notes: [
"Additional, project-based \"experience\": the portfolio's most recent Human Mode revision presented DeepOps, Orchestra, PromptOS and the ESP32-CAM Vision Suite as build-out work alongside formal roles. Those are recorded under Projects rather than employment, because the portfolio does not document them as employment. Independent project work, open-source contribution, and this portfolio's build history are treated as engineering evidence, not as employment. ",
],
}
Projects
#! Section 06 · Projects — 11 records
# ==========================================================================
# Projects
# ==========================================================================
PROJECTS: Record[str, Any] = {
section: "Section 06 · Projects",
title: "Projects",
intro: "Ordered by evidence status: flagship records first, then supporting projects, then additional projects (AgentOS public, two private repositories — access on request). Each project appears in full once, with cross-category tags at the top of its entry and an evidence status label attached to the record.",
categories: {
agentic_ai: [
],
llm_ai_infrastructure: [
],
autonomous_operations: [
],
computer_use: [
],
computer_vision_edge_ai: [
],
embedded_physical_systems: [
],
developer_open_source: [
],
},
project_details: {},
entries: [
{
anchor: "project-opencode-cua",
evidence_level: "public_repo",
tags: "Agentic AI · Computer Use · Computer Vision / Perception",
name: "OpenCode CUA Platform",
purpose: "A Computer-Use Agent platform that lets an unmodified OpenCode instance control your desktop GUI and browser through MCP servers.",
problem_solved: "AI agents are normally confined to a terminal; driving real desktop and browser software requires perception, action and — critically — recovery when an action fails.",
what_sneh_built: "A platform combining screen capture, macOS Accessibility APIs and browser automation behind MCP servers, plus dedicated verification and recovery engines so the agent can confirm and repair failed actions.",
primary_technical_evidence: "Architecture described in the portfolio and in his writing (\"the hardest part of AI agents isn't the model — it's verification and recovery when actions fail\"); repository authored by Sneh30.",
documented_result: "No usage metric published. Note: this is the same codebase family as the site's \"OpenCode\" skill entry.",
technologies_human_mode: "Python · MCP · macOS · React",
},
{
anchor: "project-webcmd",
evidence_level: "external_validation",
tags: "Developer / Open Source · external validation",
name: "webcmd — merged contribution verified",
purpose: "Contributed to agentrhq/webcmd (2,620 stars): unified output-format parsing across all builtin commands.",
problem_solved: "Every builtin command handled
-f/--format differently — an inconsistency that grows with every new command.",what_sneh_built: "A single source of truth for output-format parsing, applied across the builtin commands (15+ fixed), with 16 new tests and documentation updates.",
primary_technical_evidence: "A merged pull request in a codebase he does not own — reviewed, tested and accepted upstream.",
documented_result: "PR #190 merged 2026-08-13 with 16 commits; upstream project 2,620 stars (GitHub API, 2026-09-30).",
technologies_human_mode: "JavaScript · Open Source · CLI",
},
{
anchor: "project-promptos",
evidence_level: "public_repo",
evidence_status: "self_reported",
tags: "LLM / AI Infrastructure",
name: "PromptOS",
purpose: "macOS desktop app that treats prompts as source code and compiles them through a multi-pass optimisation pipeline.",
problem_solved: "Raw prompts are expensive and inconsistent; PromptOS rewrites and compresses them before they reach the model, locally or via cloud providers.",
what_sneh_built: "A Rust core with custom lexer, parser and AST and ten optimisation passes across four compilation modes; WASM plugin support; local inference via Llama.cpp on Metal GPU; a native SwiftUI shell.",
primary_technical_evidence: "Compiler-style architecture (not string templating), self-contained desktop distribution, and local-accelerated inference.",
documented_result: "Résumé: up to 80–85% lower inference cost in Economy mode via multi-pass token compression, tested against frontier LLM pricing. source: résumé PDF",
technologies_human_mode: "Rust · Swift · Llama.cpp · WASM · Metal GPU",
},
{
anchor: "project-orchestra",
evidence_level: "public_repo",
tags: "Agentic AI · LLM / AI Infrastructure · Autonomous Operations",
name: "Orchestra",
purpose: "Production-oriented AI research orchestration with 5 specialised LangGraph agents producing verified reports with confidence scoring.",
problem_solved: "Automated research that is trustworthy: search, extract, verify, synthesize and critique must be separated so claims can be checked rather than confidently hallucinated.",
what_sneh_built: "The five-agent LangGraph pipeline, FastAPI services, PostgreSQL persistence, Prometheus instrumentation, a test suite with a deterministic CI mode, and Docker Compose packaging for one-command deployment.",
primary_technical_evidence: "Agent decomposition with a verification stage, confidence scoring, observability, and CI-reproducible behaviour — the parts that separate a demo from a system.",
documented_result: "5 cooperating agents · verified reports with confidence scoring · one-command deployment. No traffic metric published.",
technologies_human_mode: "Python · LangGraph · FastAPI · PostgreSQL · Prometheus",
},
{
anchor: "project-deepops",
evidence_level: "public_repo",
evidence_status: "experimental",
tags: "AI Infrastructure · Autonomous Operations",
name: "DeepOps",
purpose: "Experimental autonomous SRE platform — active development: interprets natural-language incident input, performs root-cause analysis, and plans remediation.",
problem_solved: "Infrastructure incident handling depends on people reading signals under pressure; DeepOps interprets natural-language commands, performs root-cause analysis and proposes/executes remediation.",
what_sneh_built: "The multi-agent operations pipeline (LangGraph, FastAPI), including cost-optimisation logic inside the remediation planner.",
primary_technical_evidence: "Natural language → incident interpretation → root-cause analysis → remediation plan, with cost considered as a first-class constraint.",
documented_result: "Résumé status: early prototype — the incident-interpretation and remediation-planning modules exist, but the end-to-end pipeline is not yet wired together. experimental — not a claimed production system",
technologies_human_mode: "Python · LangGraph · FastAPI",
},
{
anchor: "project-esp32",
evidence_level: "public_repo",
tags: "Computer Vision / Edge AI · Embedded / Physical Systems",
name: "ESP32-CAM Vision Suite",
purpose: "Three ESP32-CAM projects: ML dataset capture, real-time colour tracking robot, and a 60 FPS MJPG streaming server.",
problem_solved: "Doing useful computer vision on hardware with 512 KB RAM, without assuming a GPU or a large model.",
what_sneh_built: "Web-based image capture for ML datasets; an autonomous colour-tracking robot with on-device RGB565→RGB888→HSV conversion and L298N motor control; an MJPEG streaming server supporting 15+ camera board variants.",
primary_technical_evidence: "Math-driven perception (HSV quantisation, centroid tracking) chosen deliberately instead of neural inference to fit the constraint.",
documented_result: "3 projects · 60 FPS streaming · 15+ board variants · 512 KB RAM ceiling.",
technologies_human_mode: "C · ESP32 · Computer Vision · HSV",
},
{
anchor: "project-catalyst",
evidence_level: "public_repo",
tags: "Developer / Full-stack",
name: "Catalyst",
purpose: "AI-powered career operating system with resume auditing, job scraping, ATS-optimized resume generation and weekly progress tracking.",
problem_solved: "Job-search work is fragmented across tools; Catalyst keeps audit → scrape → generate → track in one SQLite-backed system.",
what_sneh_built: "Backend and application logic (FastAPI/Python), the resume-audit and ATS-generation flows, job scraping, and the tracker.",
primary_technical_evidence: "End-to-end product flow with persistence, per the repository description (\"resume audit, job scraping, ATS resume generation, content engine, weekly tracker\").",
documented_result: "No usage metrics published.",
technologies_human_mode: "Python · FastAPI · JavaScript · SQLite",
},
{
anchor: "project-portfolio",
evidence_level: "public_repo",
tags: "Developer / Open Source",
name: "Portfolio (this site)",
purpose: "The site you are reading: semantic HTML/CSS/JS with canvas rendering, scroll-driven animation, and live GitHub/LeetCode API integration.",
problem_solved: "Present engineering work credibly, in two modes: a designed experience for humans and a structured record for automated parsing.",
what_sneh_built: "Front-end, canvas terrain renderer, GSAP scroll choreography, live API integrations, and this Agent Mode record.",
primary_technical_evidence: "Open repository with full history; no framework dependency for the agent-facing layer.",
technologies_human_mode: "HTML · CSS · JavaScript",
},
],
additional_projects: [
{
anchor: "project-agentos",
evidence_level: "public_repo",
tags: "Agentic AI",
name: "AgentOS public_repo",
purpose: "A personal agent operating system with 6 composable layers and 250+ app integrations.",
problem_solved: "Personal AI tooling is fragmented; AgentOS proposes one layered runtime (Interface, Model Router, Simulation Engine, Memory Graph, Guard Layer, Execution Plane).",
what_sneh_built: "Per the portfolio: the six-layer architecture with model routing, memory, guardrails and an execution plane wired to 250+ app integrations via Composio.",
primary_technical_evidence: "Public repository at github.com/Sneh30/AgentOS (TypeScript). Its README documents the same six layers (Interface, Model Router, Simulation Engine, Memory Graph, Guard Layer, Execution Plane) and the Composio execution plane (checked 2026-10-20). Documented result6 layers · 250+ app integrations — both recorded in the repository README (Execution Plane: Composio, 250+ apps). The integration count is the repository’s own figure, not independently counted (checked 2026-10-20). Technologies (Human Mode)TypeScript · Next.js · Supabase · Composio VerificationPublic repository verified 2026-10-20. github.com/Sneh30/AgentOS resolves anonymously (HTTP 200) and its README matches the architecture and integration claims above. The 250+ figure is stated by the repository rather than measured independently. The separate public
composio repository under the same account is a different project (an API/MCP research agent). DemonstratesAgentic AI (public code) ",},
{
anchor: "project-agentic-hinge",
evidence_level: "private_available_on_request",
tags: "Computer Use · Computer Vision · Agentic AI",
name: "Mobile GUI Automation Agent private_available_on_request",
purpose: "Computer-use automation over iPhone Mirroring combining screen perception, computer vision, and GUI automation. The original prototype drove a dating app (Hinge) interface to generate personalised reply suggestions.",
problem_solved: "Automating a mobile-only interface from the desktop requires bridging mirroring, screen perception and GUI control.",
what_sneh_built: "Per the portfolio: a macOS agent combining computer vision with GUI automation over iPhone Mirroring.",
primary_technical_evidence: "Private repository at github.com/Sneh30/Agentic-hinge (Swift). Its README matches this description — a macOS GUI agent that automates Hinge over iPhone Mirroring, published with an explainer video (checked 2026-10-20). Anonymous visitors receive a 404 because the repository is private; access is available on request. Documented resultNo numeric metric published. Technologies (Human Mode)Swift · Computer Vision · GUI Automation VerificationRepository exists but is private (checked 2026-10-20). The repository matches the description above, so the project is real rather than a missing link; it cannot be inspected anonymously, so its claims are not publicly verifiable. Access is available on request. DemonstratesComputer use · vision-driven automation (private repository) ",
},
{
anchor: "project-pocket-typist",
evidence_level: "private_available_on_request",
tags: "Embedded / Physical Systems",
name: "Pocket Typist private_available_on_request",
purpose: "ESP32 BLE HID firmware with a Cloudflare Workers relay for remote text-input experiments: text pushed to the device over HTTP and typed as a Bluetooth keyboard.",
problem_solved: "Getting text from any device onto any machine without installing software — solved in firmware rather than in an app.",
what_sneh_built: "Per the portfolio: ESP32 BLE HID keyboard firmware plus a Cloudflare Workers backend that pushes text to the device.",
primary_technical_evidence: "Private repository at github.com/Sneh30/pocket-typist (C++). Its README matches this description — ESP32-CAM BLE HID keyboard firmware that types text relayed by a Cloudflare Worker with a Durable Object (checked 2026-10-20). Anonymous visitors receive a 404 because the repository is private; access is available on request. Documented resultNo numeric metric published. Technologies (Human Mode)C++ · ESP32 · BLE · Cloudflare VerificationRepository exists but is private (checked 2026-10-20). The repository matches the description above, so the project is real rather than a missing link; it cannot be inspected anonymously, so its claims are not publicly verifiable. Access is available on request. DemonstratesEmbedded firmware + cloud integration (private repository) ",
},
],
additional_projects_note: "The three projects below are described from the portfolio's own text; their repository links now resolve (checked 2026-10-20). AgentOS is public and independently checkable, while Agentic-hinge and Pocket Typist are private repositories — real and inspectable by the owner, but not publicly verifiable, so access is available on request.",
}
External Evidence
#! Section 07 · External validation — PR #190
# ==========================================================================
# External Evidence
# ==========================================================================
EXTERNAL_EVIDENCE: Record[str, Any] = {
section: "Section 07 · External validation",
title: "External Evidence",
intro: "Externally verifiable contributions: code reviewed and merged into a project the candidate does not own, plus public activity records.",
entries: [
{
anchor: "evidence-webcmd-pr",
tags: "Merged pull request · independently verifiable",
name: "PR #190 — Normalize output-format parsing across builtin commands",
repository: "agentrhq/webcmd — open-source browser-automation / CLI toolkit for AI agents, 2,620 stars (GitHub API, 2026-09-30). The portfolio previously described it as a 1.6k-star project; the figure above is the current verified count.",
contribution: "Unified
-f/--format parsing so every builtin command accepts the same output formats — table, plain, json, yaml, md, csv — by creating a single source of truth.",what_changed: "15+ commands fixed, 16 tests added, documentation updated.",
scale_status: "Merged 2026-08-13, containing 16 commits authored by Sneh30 (verified via GitHub API, 2026-09-30).",
context: "It shows the ability to enter an unfamiliar, popular codebase, make a cross-cutting refactor safely (single source of truth + tests), and land it through review — the exact behaviour expected of a professional engineer on day one.",
related_evidence: "Personal fork Sneh30/webcmd · portfolio write-up \"What I Learned Contributing to a 2.6k-Star Project\" (Aug 2026) · upstream repository.",
},
],
cards: [
{
label: "GitHub account",
title: "Public activity",
points: [
"13 public repositories spanning AI agents, LLM systems, infrastructure, embedded systems, and developer tooling (GitHub API, 2026-09-30)",
"732+ contributions shown on the portfolio self-published figure",
],
},
{
label: "Writing",
title: "Build-in-the-open",
points: [
"6 dated technical posts in the portfolio (May–Aug 2026) covering computer-use agents, open-source contribution, prompt compilation, embedded CV, GUI automation and autonomous infrastructure.",
],
},
],
notes: [
"Transparency: no open-source claim on this page goes beyond the merged PR documented above. The portfolio does not claim maintainership, funding, or adoption of any third-party project, and it does not claim stars on his own repositories (his own repos currently show 0 stars — the 2.6k figure belongs to the upstream project he contributed to).",
"Additional profiles: LeetCode sneh_sinha3 — 37 problems solved (32 easy / 5 medium / 0 hard), ranking 3,387,548 (LeetCode API, 2026-09-30), loaded live by Human Mode. Supporting signal only; not a headline qualification.",
],
}
Technical Skills verbatim from Human Mode
{
"@file": "SKILLS",
"@section": "Section 08 · Technical skills (verbatim from Human Mode)",
"@title": "Technical Skills verbatim from Human Mode",
"@generated from": "structured candidate record, Human Mode sources",
"section": "Section 08 · Skills",
"title": "Technical Skills verbatim from Human Mode",
"intro": "Reproduced exactly as listed in the Human Mode skills section. This list has not been expanded, upgraded, or reinterpreted for Agent Mode. Additional technologies appear only as project tags attached to specific projects, and are labelled as such.",
"cards": [
{
"title": "Languages",
"points": [
"Python",
"Java",
"SQL"
]
},
{
"title": "AI / Agent Systems",
"points": [
"LangGraph",
"OpenAI Codex",
"Claude Code",
"OpenCode",
"OpenRouter",
"Ollama",
"Llama.cpp"
]
},
{
"title": "AI / ML",
"points": [
"Hugging Face Transformers",
"LLM Inference",
"Multi-Agent Systems",
"Prompt Engineering",
"Edge AI"
]
},
{
"title": "Backend & Cloud",
"points": [
"FastAPI",
"PostgreSQL",
"Prometheus",
"REST APIs",
"AWS",
"Observability",
"Incident Response"
]
},
{
"title": "Infra & Tooling",
"points": [
"Docker",
"Docker Compose",
"Git",
"GitHub Actions",
"CI/CD",
"Linux",
"WASM",
"Metal (GPU)"
]
},
{
"title": "Embedded & Hardware",
"points": [
"ESP32",
"Raspberry Pi",
"Edge Impulse"
]
}
],
"notes": [
"Also stated in Human Mode (project tags, not skill claims): Swift, TypeScript, C, C++, JavaScript, React, Next.js, SQLite, MCP, Cloudflare, WASM, Computer Vision, BLE — each appears on a specific project card in Human Mode and is claimed here only in that project's context. Exploring (per Human Mode): Multi-Agent Systems, Edge AI. The portfolio does not claim production experience in Kubernetes, and Kubernetes is not listed in the Human Mode skills section. "
]
}
Where This Evidence Applies
# Section 09 · where_this_evidence_applies
# Where This Evidence Applies
# source: structured candidate record · record/09_roles.yaml
"section": "Section 09 · where_this_evidence_applies"
"title": "Where This Evidence Applies"
"intro": "role → relevant capabilities → evidence. Factual mapping for matching purposes; no scores, rankings, or fit percentages."
"roles":
- "role": "AI Engineer"
"Role ID": "role-ai-engineer"
"Support": "strongly_supported"
"Relevant capabilities": "Multiple built AI systems spanning agents, multi-agent pipelines, and LLM tooling; portfolio focus explicitly on \"AI Agents, LLM Cost & Infra\"."
"Context": "The evidence covers complete systems rather than isolated model calls — perception, orchestration, cost control and packaging — across AI agents, multi-agent pipelines and LLM tooling, each in a published artifact."
"Technical topics": "Verification and recovery in the computer-use agent; Orchestra's verify stage and deterministic CI mode; how PromptOS measures its cost savings."
- "role": "Agentic AI Engineer"
"Role ID": "role-agentic-ai-engineer"
"Support": "strongly_supported"
"Relevant capabilities": "A computer-use agent with verification/recovery, a five-agent research system with an explicit verify stage, and an agent operating-system architecture."
"Evidence · projects": "OpenCode CUA Platform · Orchestra · AgentOS (public repo) · Mobile GUI Automation Agent (private repo)"
"Skills": "LangGraph · OpenCode · OpenRouter · Ollama · Llama.cpp · Multi-Agent Systems"
"Context": "Agentic products fail on reliability rather than on model choice. The evidence here centres on failure handling, explicit verification stages and tool integration — the parts that decide whether an agent works outside a demo."
"Technical topics": "MCP server design trade-offs; confidence scoring and disagreement handling between agents; guardrails and execution boundaries."
- "role": "AI Infrastructure Engineer"
"Role ID": "role-ai-infrastructure-engineer"
"Support": "supported"
"Relevant capabilities": "DeepOps (AI-driven operations) plus a documented 400+ node operations engagement with a ~40% improvement in incidents and recovery time."
"Context": "The combination of building AI that operates infrastructure (DeepOps) and operating real infrastructure (400+ node engagement) means the automation work is informed by actual incident behaviour rather than a sandbox."
"Technical topics": "What changed to produce the ~40% improvement and how it was measured; DeepOps remediation-planner design; what unattended operation would require."
- "role": "LLM Engineer"
"Role ID": "role-llm-engineer"
"Support": "supported"
"Relevant capabilities": "Prompt compilation with measured cost reduction; local inference on Metal; provider abstraction; day-to-day agent tooling."
"Caveat": "The portfolio does not claim model training, fine-tuning, or research; the LLM evidence is systems- and cost-oriented."
"Context": "PromptOS is a systems answer to cost and latency discipline: a compiler pipeline plus a résumé-reported cost figure (up to 80–85% in Economy mode, self-reported)."
"Technical topics": "The ten optimisation passes and which contributes most; guarding output quality under compression; how the 80–85% figure was measured."
- "role": "Autonomous Systems Engineer"
"Role ID": "role-autonomous-systems-engineer"
"Support": "supported"
"Relevant capabilities": "Self-driving perception/control in software (colour-tracking robot), safety-critical braking firmware on a physical vehicle, and autonomous remediation in infrastructure."
"Evidence · projects / experience": "ESP32-CAM Vision Suite · Team Stratos (ABS firmware) · DeepOps"
"Context": "Autonomous systems span software control loops and physical hardware; this profile has both — on-device perception and motor control, safety-critical braking firmware, and autonomous remediation logic."
"Technical topics": "ABS control logic and track-test debugging; the colour-tracking robot's control loop; failure handling in DeepOps remediation."
- "role": "AI/Software Systems Engineer"
"Role ID": "role-ai-software-systems-engineer"
"Support": "supported"
"Relevant capabilities": "Backend services, databases, observability, CI/CD, desktop applications in Rust/Swift, and embedded firmware — a systems profile rather than a notebook profile."
"Context": "Systems roles reward breadth across services, storage, CI and delivery; the evidence includes backends with persistence and observability, desktop applications, embedded firmware, and a deployed site."
"Technical topics": "PostgreSQL and Prometheus choices in Orchestra; PromptOS packaging and distribution; the test strategy behind the webcmd refactor."
- "role": "ML Engineer"
"Role ID": "role-ml-engineer"
"Support": "partially_supported"
"Relevant capabilities": "Hugging Face Transformers and Edge AI in the skills list; ML dataset capture tooling; classical CV on-device (HSV quantisation, centroid tracking); Edge Impulse."
"Caveat": "The portfolio does not document model training, evaluation benchmarks, or deployed ML services. Human Mode lists Multi-Agent Systems and Edge AI under \"Exploring\". This is adjacent evidence, not a primary claim."
"Context": "Adjacent evidence only: on-device classical CV, dataset-capture tooling and Edge Impulse appear in the portfolio, but model training and evaluation are not documented."
"Technical topics": "The dataset-capture workflow; when classical CV beats a model on-device; what an Edge Impulse deployment would look like here."
- "role": "Software Engineer"
"Role ID": "role-software-engineer"
"Support": "supported"
"Relevant capabilities": "Backend and full-stack delivery with persistence and APIs, plus code that passed review in a project the candidate does not own."
"Context": "Generalist software evidence rather than one framework: Python services, a JavaScript/SQLite product, a deployed site with API integrations, and merged code in an externally maintained project."
"Technical topics": "Schema and migration choices in Catalyst; the test strategy behind the webcmd refactor; how the portfolio site is structured for parsers."
- "role": "Backend Engineer"
"Role ID": "role-backend-engineer"
"Support": "supported"
"Relevant capabilities": "Complete applications with persistence, REST APIs, and integrations — the \"Backend & Full-Stack Delivery\" capability."
"Skills (Human Mode)": "Python · FastAPI · PostgreSQL · SQL · REST APIs · Prometheus"
"Context": "Both projects expose HTTP services backed by a real store and ship observability hooks; no usage metrics are published for either."
"Technical topics": "Schema and migration choices in Catalyst; Prometheus metric design in Orchestra; where the API boundaries sit."
- "role": "Python Developer"
"Role ID": "role-python-developer"
"Support": "supported"
"Basis": "Python listed first under Languages in Human Mode; all four projects list Python in Human Mode"
"Relevant capabilities": "Python is listed first under Languages in Human Mode, and four published projects are Python end to end."
"Skills (Human Mode)": "Python · FastAPI · LangGraph · PostgreSQL · Docker"
"Context": "The same language appears in an agent control loop, a multi-agent research system, an autonomous operations prototype, and a product backend — different problem shapes, one toolchain."
"Technical topics": "Where async and concurrency matter in Orchestra; keeping the agent loop in OpenCode CUA deterministic; FastAPI service structure shared by two projects."
- "role": "DevOps / SRE Engineer"
"Role ID": "role-devops-sre-engineer"
"Support": "supported"
"Basis": "project + engagement"
"Relevant capabilities": "AI Infrastructure & Autonomous Operations plus a documented 400+ node operations engagement with a ~40% improvement in incidents and recovery time."
"Caveat": "No Kubernetes experience is claimed — Kubernetes is not listed in Human Mode skills. The evidenced tooling is Docker, Docker Compose, GitHub Actions CI/CD, Linux, Prometheus, AWS, and incident response."
"Context": "The operations engagement is real and dated; the automation side is a prototype, so the two are presented separately rather than as a production SRE track record."
"Technical topics": "What produced the ~40% improvement and how it was measured; alerting and dashboard choices in Orchestra; what unattended remediation would require."
- "role": "Automation Engineer"
"Role ID": "role-automation-engineer"
"Support": "supported"
"Relevant capabilities": "Autonomous remediation in infrastructure, command behaviour in an upstream CLI tool, and screen-driven GUI automation."
"Evidence · projects": "DeepOps · webcmd (merged contribution) · Mobile GUI Automation Agent private_available_on_request"
"Caveat": "The automation evidence is infrastructure remediation, CLI behaviour, and desktop/mobile GUI control. No RPA platform or test-automation framework is claimed."
"Context": "Automation here means code that acts on its own signal — incident input to remediation plans, 15+ CLI commands fixed and covered by tests, and GUI control driven by screen perception."
"Technical topics": "Guardrails around automatic remediation; how the GUI agent decides when to act; test coverage for automation paths."
- "role": "Embedded Systems Engineer"
"Role ID": "role-embedded-systems-engineer"
"Support": "supported"
"Relevant capabilities": "Computer vision and firmware on constrained hardware, plus a safety-critical subsystem owned through physical validation."
"Skills (Human Mode)": "ESP32 · Raspberry Pi · Edge Impulse"
"Context": "Evidence runs from a 512 KB vision pipeline and a 60 FPS streaming server to BLE HID firmware and ABS braking firmware deployed on a vehicle."
"Technical topics": "Memory budgets in the vision pipeline; BLE HID timing and the Cloudflare relay; how ABS firmware was validated on track."
- "role": "Robotics Engineer"
"Role ID": "role-robotics-engineer"
"Support": "supported"
"Relevant capabilities": "On-device perception and control, safety-critical vehicle firmware, and a B.E. in Robotics and Artificial Intelligence (2023–2027)."
"Caveat": "No ROS, manipulator, or control-system tooling is listed in Human Mode skills. The robotics evidence is ESP32 perception and an autonomous colour-tracking robot, BLE device firmware, and safety-critical vehicle firmware, alongside the B.E. in Robotics and Artificial Intelligence."
"Context": "Robotics evidence is perception, actuation and firmware on real hardware: an autonomous colour-tracking robot, BLE device firmware, and braking firmware on a competition vehicle."
"Technical topics": "Control loop and failure behaviour of the colour-tracking robot; sensor and actuator choices under a 512 KB budget; what the ABS firmware does on a lost wheel-speed signal."
- "role": "Computer Vision Engineer"
"Role ID": "role-computer-vision-engineer"
"Support": "supported"
"Relevant capabilities": "Edge computer vision on constrained hardware and screen perception driving GUI automation."
"Evidence · projects": "ESP32-CAM Vision Suite · Mobile GUI Automation Agent private_available_on_request"
"Skills (Human Mode)": "Edge Impulse · Edge AI"
"Caveat": "No large-scale dataset, benchmark, or published model weights are claimed; the vision work is edge and screen-perception driven."
"Context": "HSV quantisation and centroid tracking on-device, ML dataset capture tooling, and perception feeding GUI control — vision chosen to fit the hardware rather than a benchmark."
"Technical topics": "When classical CV beats a model on-device; perception latency inside the GUI agent; dataset capture quality."
Record Summary
# Record Summary
section: Section 10 · Summary
title: Record Summary
## Recorded work
Built systems and prototypes across AI agents, LLM systems, infrastructure, and embedded/physical-world computing: a computer-use agent (OpenCode CUA), multi-agent research orchestration with verification (Orchestra), experimental infrastructure-operations AI (DeepOps), prompt compilation with a candidate-reported cost reduction (PromptOS), embedded vision inside 512 KB RAM (ESP32-CAM Vision Suite), and safety-critical vehicle firmware (Team Stratos ABS).
## Verification basis
- Externally verifiable code: a merged 16-commit PR into a 2.6k-star repository he does not own, plus live repositories for eight of the eleven listed projects.
- Documented constraint sets: a 400+ node network with a stated ~40% improvement, a 50-member team, safety-critical braking firmware, and a 512 KB memory ceiling.
- Consistency across sources: the Human Mode portfolio, the résumé PDF, the GitHub account, and this record carry the same facts and numbers.
- Documented limitations: where evidence is missing (unavailable repositories, undisclosed employer, self-published contribution counts), this page says so instead of implying more.
Technical Topics
# Technical Topics
section: Section 11 · Technical topics
title: Technical Topics
intro: Topic index derived from the published work — discussion areas that follow from it, listed for reference.
## topics
### Computer-use agent (OpenCode CUA)
1. How does the system decide that an action "succeeded"? What does verification actually check?
2. Describe a failure the recovery engine handles. What is the retry/backoff policy?
3. Why MCP servers — what broke when you tried a tighter integration?
4. How do Accessibility API permissions, screen latency and focus stealing affect reliability?
### Multi-agent architecture (Orchestra)
1. Why five agents rather than one agent with five tools? Where is the boundary drawn?
2. How is confidence scoring computed, and what happens when verify disagrees with extract?
3. What does the deterministic CI mode freeze — model outputs, tool calls, or ordering?
4. What are the latency and cost characteristics of a full research run?
### Infrastructure & operations
1. In the 400+ node engagement, what specifically changed to cut incidents and recovery time by ~40%?
2. How was the baseline ("prior comparable period") defined and measured?
3. Walk through one incident you handled end to end. What was your first signal?
4. How did you decide what to automate versus document?
### LLM systems & cost (PromptOS)
1. What are the ten passes, and which one contributes the most savings?
2. How do you prevent compression from degrading model output quality?
3. How was the 80–85% Economy-mode figure measured against frontier pricing?
4. Why a Rust compiler core instead of a Python pipeline?
### Embedded & safety-critical
1. How does the ABS firmware decide to release brake pressure? What sensors and timings are involved?
2. What failed during track testing, and how did you debug it without a debugger on the vehicle?
3. How do you fit vision processing into 512 KB RAM — what did you give up?
4. What would change if the colour-tracking robot had to run unattended for an hour?
### Engineering judgement & open source
1. Walk through PR #190: how did you introduce a single source of truth without breaking 15+ commands?
2. What tradeoffs did you make in DeepOps, and what would a production implementation need that it lacks today?
3. What did you personally implement in Team Stratos versus delegate to the team?
4. Of everything here, what would you rebuild differently now — and why?
Evidence of Capability
{
"@file": "EVIDENCE",
"@section": "Section 12 · Evidence graph — 9 capabilities",
"@title": "Evidence of Capability",
"@generated from": "structured candidate record, Human Mode sources",
"section": "Section 12 · Evidence graph",
"title": "Evidence of Capability",
"intro": "Every major claim follows the same chain: capability → what was built → what Sneh personally did → result/scale → technologies already present in Human Mode → verification link. This is the evidence graph, walkable in either direction by a human reader or an automated parser. Each block carries a qualitative evidence status label (`public_repo` · `private_available_on_request` · `external_validation` · `documented_experience` · `self_reported` · `experimental`); no numeric scores are assigned.",
"capabilities": [
{
"anchor": "capability-agents",
"evidence level": "public_repo",
"capability": "AI Agents & Computer Use capability 01 public_repo",
"summary": "Primary evidence: a platform that lets an AI agent operate a real desktop and browser, including failure handling.",
"What was built": "OpenCode CUA Platform — a Computer-Use Agent platform that lets an unmodified OpenCode instance control the desktop GUI and browser through MCP servers.",
"What Sneh personally did": "Author of the repository (Sneh30/opencode-cua-platform). The portfolio describes the architecture he designed: screen capture, macOS Accessibility APIs, browser automation, plus verification and recovery engines for failed actions.",
"Result / scale": "No numeric usage metric is documented. The documented result is architectural: an end-to-end computer-use loop with explicit verification and recovery stages rather than blind action execution.",
"Technologies (Human Mode)": "Python, MCP, macOS, React · skills: OpenCode, LangGraph, LLM Inference, Multi-Agent Systems.",
"Demonstrates": "Agentic AI · Computer use · Tool/middleware integration (MCP) · Reliability engineering (verification & recovery)."
},
{
"anchor": "capability-multagent",
"evidence level": "public_repo",
"capability": "Multi-Agent Orchestration capability 02 public_repo",
"summary": "Primary evidence: five specialised agents with verification, persistence, observability and CI.",
"What was built": "Orchestra — production-oriented AI research orchestration with 5 specialised LangGraph agents (search, extract, verify, synthesize, critique) producing verified reports with confidence scoring.",
"What Sneh personally did": "Designed and built the multi-agent pipeline (repository owner: Sneh30/Orchestra). The résumé documents the parts usually skipped: a full test suite with a deterministic CI mode, PostgreSQL persistence, Prometheus instrumentation, and Docker Compose packaging for one-command deployment.",
"Result / scale": "5 cooperating agents; verified-report output with confidence scoring; reproducible one-command deployment. No user/traffic metric is documented.",
"Technologies (Human Mode)": "Python, LangGraph, FastAPI, PostgreSQL, Prometheus.",
"Demonstrates": "Multi-agent systems · Backend engineering · Testing & CI discipline · Observability."
},
{
"anchor": "capability-ai-infra",
"evidence level": "public_repo",
"capability": "AI Infrastructure & Autonomous Operations capability 03 public_repo experimental",
"summary": "Primary evidence: AI that takes natural-language incidents to root cause and remediation — plus real operations work on 400+ nodes.",
"What was built": "DeepOps — experimental autonomous SRE platform, active development: natural-language infrastructure commands, anomaly/root-cause analysis and proposed remediation.",
"What Sneh personally did": "Creator and lead developer (repository owner: Sneh30/deepops). The résumé documents cost-optimisation logic built into the remediation planner.",
"Result / scale": "Documented status: early prototype — the incident-interpretation and remediation-planning modules exist, but the end-to-end pipeline is not yet wired together. experimental, not a claimed production deployment",
"Technologies (Human Mode)": "Python, LangGraph, FastAPI · skills: Prometheus, Observability, Incident Response, AWS, Docker.",
"Demonstrates": "AI infrastructure tooling (experimental) · Incident-response reasoning · Cost-aware systems design."
},
{
"anchor": "capability-llm",
"evidence level": "public_repo",
"capability": "LLM Systems & Cost/Latency Engineering capability 04 public_repo self-reported metric",
"summary": "Primary evidence: a real compiler pipeline for prompts, with local inference and measured cost reduction.",
"What was built": "PromptOS — a macOS desktop app that treats prompts as source code: a Rust core with custom lexer, parser and AST, ten optimisation passes across four compilation modes, WASM plugins, local LLM inference via Llama.cpp/Metal GPU, and cloud provider support.",
"What Sneh personally did": "Creator and lead developer (repository owner: Sneh30/promptos); built as a self-contained macOS app with a native SwiftUI shell.",
"Result / scale": "Résumé: reduced inference cost by up to 80–85% in Economy mode via multi-pass token compression, tested against frontier LLM pricing. source: résumé PDF linked from this portfolio",
"Technologies (Human Mode)": "Rust, Swift, Llama.cpp, WASM, Metal GPU · skills: LLM Inference, Prompt Engineering.",
"Demonstrates": "LLM internals awareness · Systems programming (Rust) · Desktop software · Quantified cost optimisation."
},
{
"anchor": "capability-ops",
"evidence level": "documented_experience",
"capability": "Real-World Infrastructure Operations capability 05 documented_experience self_reported",
"summary": "Primary evidence: an on-site, time-boxed engagement with a documented operational outcome.",
"What was built / done": "Strengthened infrastructure operations for a 400+ node exam-facility network during a one-month on-site engagement (Aug 2026).",
"What Sneh personally did": "Recruited for and delivered the engagement independently (\"Independent Engagement\" per the portfolio). Exact tooling and scope per shift are not detailed in the portfolio.",
"Result / scale": "Documented: cut infrastructure incidents and average recovery time by ~40% versus the prior comparable period (the site notes a prior period of over 5 days).",
"Technologies (Human Mode)": "Skills: Linux, AWS, Docker, CI/CD, Observability, Incident Response, Prometheus.",
"Verification": "Portfolio experience section · résumé (PDF) employer name and raw before/after data not published",
"Demonstrates": "Live operations · Incident response · Measurable reliability improvement.",
"verification limit": "Verification limit: the portfolio does not currently provide sufficient evidence to verify this claim beyond the candidate's own published statement — no employer name, ticket data, or third-party reference is published. The before/after incident dataset is not published."
},
{
"anchor": "capability-embedded",
"evidence level": "public_repo",
"capability": "Embedded Systems & Edge Computer Vision capability 06 public_repo",
"summary": "Primary evidence: three working vision systems on constrained hardware, plus safety-critical vehicle firmware.",
"What was built": "ESP32-CAM Vision Suite — three projects: web-based image capture for ML datasets, a real-time colour-tracking robot, and a full MJPG video streaming server at 60 FPS.",
"What Sneh personally did": "Repository owner (Sneh30/esp32-cam-vision). The résumé details the engineering: math-driven approaches (HSV quantisation, centroid tracking) instead of neural inference, on-device RGB565→RGB888→HSV conversion, L298N motor control, an image-collection server, and streaming supporting 15+ camera board variants — all within 512 KB RAM.",
"Result / scale": "3 projects · 60 FPS streaming · 15+ supported camera board variants · 512 KB RAM ceiling.",
"Technologies (Human Mode)": "C, ESP32, Computer Vision, HSV · skills: Edge Impulse, Raspberry Pi, Edge AI.",
"Demonstrates": "Edge AI · Real-time CV · Resource-constrained design · Hardware/software integration."
},
{
"anchor": "capability-safety",
"evidence level": "documented_experience",
"capability": "Safety-Critical Systems Engineering capability 07 documented_experience team project",
"summary": "Primary evidence: ownership of the most safety-critical subsystem on a real competition vehicle.",
"What was built / done": "Developed and deployed the ABS (Anti-lock Braking System) firmware for a 50-member team's competition vehicle — described in the portfolio as the most safety-critical subsystem.",
"What Sneh personally did": "Software & Electronics Lead for eleven months (Jul 2025 – May 2026): owned software and electronics integration for the vehicle, including the ABS firmware.",
"Result / scale": "Validated through track testing, SAE India scrutineering and competition. 50-member team · 11-month tenure.",
"Technologies (Human Mode)": "Embedded stack as listed: ESP32, Raspberry Pi, Edge Impulse; languages C/C++ appear in project tags.",
"Verification": "Portfolio experience section · résumé (PDF) team project; individual code not separately published",
"Demonstrates": "Safety-critical firmware · Systems integration · Leadership in a hardware team · Physical-world validation."
},
{
"anchor": "capability-oss",
"evidence level": "external_validation",
"capability": "Software Engineering & Open-Source Contribution capability 08 external_validation",
"summary": "Primary evidence: code reviewed and merged into a project he does not own.",
"What was built / done": "Merged PR #190 into
agentrhq/webcmd: unified -f/--format parsing so every builtin command accepts the same output formats (table, plain, json, yaml, md, csv), created a single source of truth, fixed 15+ commands, added 16 tests, updated docs.","What Sneh personally did": "Authored all 16 commits in the merged PR (verified via the GitHub API on 2026-09-30).",
"Result / scale": "PR merged 2026-08-13 into a project with 2,620 stars (upstream
agentrhq/webcmd, GitHub API, 2026-09-30).","Technologies (Human Mode)": "JavaScript, Open Source, CLI (project tags) · skills: Git, GitHub Actions, CI/CD.",
"Demonstrates": "Reading and modifying an unfamiliar codebase · Test discipline · Review-grade communication."
},
{
"anchor": "capability-backend",
"evidence level": "public_repo",
"capability": "Backend & Full-Stack Delivery capability 09 public_repo",
"summary": "Supporting evidence: two complete product-shaped applications with persistence and APIs.",
"What was built": "Catalyst — an AI-powered career operating system with resume auditing, job scraping, ATS-optimized resume generation and weekly progress tracking stored in SQLite; and the Portfolio project itself (this site), with live GitHub/LeetCode API integration.",
"Result / scale": "Working end-to-end features (scraping → audit → generated output → tracking). No traffic or user metrics are documented.",
"Technologies (Human Mode)": "Python, FastAPI, JavaScript, SQLite, HTML/CSS/JavaScript.",
"Demonstrates": "End-to-end product delivery · Data persistence · API integration."
}
]
}
Candidate Record
{
"@file": "CANDIDATE",
"@section": "Section 13 · Structured candidate record",
"@title": "Candidate Record",
"@generated from": "structured candidate record, Human Mode sources",
"section": "Section 13 · Structured record",
"title": "Candidate Record",
"intro": "Complete, machine-readable record in semantic HTML. The same information is published as structured files: candidate.json, capabilities.json, evidence.json, agent-card.json, and llms.txt.",
"record": [
{
"Identity & contact": [
"Name: Sneh Sinha (Sneh Raj Sinha)",
"Role: AI Engineer",
"Location: Bangalore, India",
"Availability: Open to work — seeking an internship or full-time engagement (intern / entry level)",
"Phone: +91 7903907151 résumé only",
"LeetCode: sneh_sinha3"
]
},
{
"Education": [
"B.E. Robotics and Artificial Intelligence — Bangalore Institute of Technology, Bangalore, Karnataka",
"Aug 2023 – Aug 2027 (graduating 2027) · GPA 8.0/10",
"Source: résumé PDF published with this portfolio; also displayed in Human Mode."
],
"Documented highlights": [
"Merged PR #190 into agentrhq/webcmd (16 commits, 2026-08-13)",
"~40% reduction in infrastructure incidents and recovery time across a 400+ node network",
"ABS firmware developed, deployed and validated through SAE India scrutineering and competition",
"732+ GitHub contributions (portfolio figure) · 13 owned public repositories",
"Up to 80–85% LLM inference cost reduction in PromptOS Economy mode (résumé)"
]
},
{
"Experience": [
"Software & Electronics Lead — Team Stratos — Jul 2025 – May 2026, 11 months · details"
],
"Open source": [
"Sneh30/webcmd — personal fork"
]
},
{
"Projects": [
"Mobile GUI Automation Agent — computer-use over iPhone Mirroring (original prototype: Agentic-hinge) · repo private_available_on_request",
"Pocket Typist — ESP32 BLE HID firmware + Cloudflare Workers relay · repo private_available_on_request"
]
},
{
"Skills (as in Human Mode)": [],
"Languages": [
"Python",
"Java",
"SQL"
],
"AI / Agent Systems": [
"LangGraph",
"OpenAI Codex",
"Claude Code",
"OpenCode",
"OpenRouter",
"Ollama",
"Llama.cpp"
],
"AI / ML": [
"Hugging Face Transformers",
"LLM Inference",
"Multi-Agent Systems",
"Prompt Engineering",
"Edge AI"
],
"Backend & Cloud": [
"FastAPI",
"PostgreSQL",
"Prometheus",
"REST APIs",
"AWS",
"Observability",
"Incident Response"
],
"Infra & Tooling": [
"Docker",
"Docker Compose",
"Git",
"GitHub Actions",
"CI/CD",
"Linux",
"WASM",
"Metal (GPU)"
],
"Embedded & Hardware": [
"ESP32",
"Raspberry Pi",
"Edge Impulse"
]
},
{
"Evidence index": [
"Verification status of every project: Projects section"
],
"Verification & limitations": [
"Employer name for the Aug 2026 engagement is not disclosed; no raw incident data is published. The portfolio does not currently provide sufficient evidence to verify this claim beyond the candidate's own statement.",
"AgentOS resolves publicly at github.com/Sneh30/AgentOS (checked 2026-10-20) and its README supports its claims; Agentic-hinge and Pocket Typist are private repositories, so anonymous visitors receive a 404 and access is available on request — their descriptions remain the owner’s own account.",
"The 732+ contributions figure is self-published on the site and was not independently verified.",
"PromptOS 80–85% and ESP32 512 KB / 15+ board figures come from the résumé PDF, not from independent measurement.",
"DeepOps is described as an early prototype (modules built, end-to-end pipeline not yet connected) — it is not presented as a production system.",
"Team Stratos is a team project; the portfolio attributes the ABS firmware to Sneh but does not publish the code.",
"All repository links on this page were re-checked 2026-10-20; star and API figures were last verified 2026-09-30."
]
}
]
}