Why I Built a Computer-Use Agent from Scratch
Building opencode-cua-platform taught me that the hardest part of AI agents isn't the model — it's verification and recovery when actions fail.
Aug 2026
AI Engineer — Bangalore, India
Hi, I'm Sneh Sinha — AI Engineer
Builds AI systems that solve real-world problems.
(scroll)
■ About
An AI engineer from Bangalore, building at the intersection of intelligence, infrastructure, and the physical world.
// profile.ts
interface Engineer {
name: "Sneh Sinha";
location: "Bangalore, India";
role: "AI Engineer";
education: "B.E. Robotics & AI, BIT (2023–2027)";
vibe: "building";
}
I'm an AI engineer from Bangalore who builds systems at the intersection of artificial intelligence, robotics, and cloud infrastructure. I believe the most interesting problems live where software meets the physical world.
Currently exploring multimodal AI agents, computer vision on edge hardware, and autonomous cloud operations. I work with Python, Java, and SQL, plus whatever else gets the job done — Rust, C++, or TypeScript.
When I'm not shipping code, you'll find me tinkering with ESP32-CAM modules, optimizing LLM inference pipelines, or reading about distributed systems.
Focus
AI Agents
LLM Systems
AI Infrastructure
Computer Vision
Embedded AI
Open Source
Real-World Engineering
■ Projects
What I've
Built
(01)
OpenCode CUA Platform
Public repoDesktop + browser controlVerification & recovery engines
A Computer-Use Agent platform that lets an unmodified OpenCode instance control your desktop GUI and browser through MCP servers. Uses screen capture, macOS Accessibility APIs, and browser automation, with structural verification and recovery mechanisms around unreliable UI actions.
▸ Explainer video
Click for sound
GitHub →
(02)
PromptOS
Public repo Self-reported metric10-pass compilerCandidate-reported 80–85% cost reduction
macOS desktop app that treats prompts as source code — a Rust compiler with a custom lexer, parser and AST, optimization passes, WASM plugins, local llama.cpp inference via the Metal GPU, provider abstraction for cloud models, and a native SwiftUI shell. The résumé reports a candidate-reported reduction in inference cost of up to 80–85% in Economy mode (self-reported, not independently verified).
▸ Explainer video
Click for sound
GitHub →
(03)
Orchestra
Public repo5 specialized agentsCI + observability
Production-oriented multi-agent research orchestration with five specialized LangGraph agents — search, extraction, verification, synthesis, critique — producing verified reports with confidence scoring, backed by PostgreSQL, Prometheus, Docker Compose, and a deterministic CI mode with structured output.
▸ Explainer video
Click for sound
GitHub →
(04)
DeepOps
Public repo Experimental · active developmentNL → root cause → remediationCost-aware repair
Experimental autonomous SRE platform — active development. Interprets natural-language incident input, performs root-cause analysis, and plans remediation with cost-aware logic. 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 deployment.
▸ Explainer video
Click for sound
GitHub →
(05)
ESP32-CAM Vision Suite
Public repo3 projects60 FPS512 KB RAM
Three ESP32-CAM projects: web-based image capture for ML datasets, a real-time color-tracking robot, and a full-featured MJPG video streaming server at 60 FPS — math-driven CV (HSV quantisation, centroid tracking) inside 512 KB of RAM, across 15+ camera board variants.
▸ Explainer video
Click for sound
GitHub →
(06)
Catalyst
Public repoResume audit → ATS resumeJob scraping + weekly tracker
An AI-powered career operating system with resume auditing, job scraping, ATS-optimized resume generation, and weekly progress tracking — all stored in SQLite.
▸ Explainer video
Click for sound
GitHub →
(07)
Portfolio
Public repoCanvas terrain rendererAgent Mode machine record
This site: semantic HTML/CSS/JS with a canvas terrain renderer, scroll choreography, live GitHub/LeetCode stats with static fallbacks, and a structured candidate profile. Deployed on GitHub Pages at sneh30.github.io/portfolio/.
GitHub →
(08)
WebCMD Merged Upstream Contribution
External validationagentrhq/webcmd · 2,620★PR #190 merged16 commits · 15+ commands · 16 tests
Unified output-format parsing across the builtin commands of an existing open-source codebase: a single source of truth for -f/--format, 15+ commands fixed, 16 tests added, docs updated — reviewed and merged upstream as PR #190 into webcmd (Aug 13, 2026). Demonstrates understanding an unfamiliar codebase, a non-trivial cross-cutting fix, automated testing, and upstream review.
(03)
Most projects are open source — check the repos for yourself.
Shipping fast, breaking things, and learning out loud.
■ Real-World Experience
Work
History
Real-world infrastructure experience and safety-critical embedded work, both in live operational environments.
Aug 2026
1 month
Infrastructure Engineer
Independent Engagement
400+ nodes
~40% incident & recovery improvement
Demonstrates the ability to work independently in a live operational environment and deliver a reported reliability improvement during a short engagement: one month on-site, strengthening infrastructure operations for a 400+ node exam-facility network, with a reported ~40% reduction in infrastructure incidents and average recovery time versus the prior comparable period (self-reported).
Jul 2025 — May 2026
11 months
Software & Electronics Lead
Team Stratos
50-member team
Safety-critical ABS firmware
SAE India validated
Software and electronics integration for a 50-member team's competition vehicle; developed and deployed its ABS (Anti-lock Braking System) firmware, the most safety-critical subsystem, validated through track testing and SAE India scrutineering and competition.
■ Technical Skills
Tech &
Tools
Languages
AI / Agent Systems
AI / ML
Backend & Cloud
Infra & Tooling
Embedded & Hardware
Languages3
FocusAI Agents, LLM Cost & Infra
ExploringMulti-Agent Systems, Edge AI
(02)
Full-stack AI engineer comfortable from embedded firmware to cloud orchestration.
The right tools are the ones that get out of your way when you're in flow.
■ Open Source
GitHub
Pulse
13
Owned Repos
732+
Contributions
5
Languages
3
Followers
Open Source Work
Merged
Normalize output-format parsing across builtin commands
Unified -f/--format parsing so every builtin command in webcmd accepts the same output formats — table, plain, json, yaml, md, csv. Created a single source of truth, fixed 15+ commands, added 16 tests.
Contribution activity (illustrative pattern)
■ Evidence & Verification
How claims are
Labeled
Every project and claim on this site carries a qualitative evidence status. No numeric evidence scores are used. Numbers attributed to the résumé are marked self-reported; dynamic statistics were last verified on 2026-09-30 against the GitHub and LeetCode APIs.
Public repo
Repository link resolves and the code is publicly readable.
External validation
Reviewed and accepted by someone outside this portfolio (e.g., a merged upstream pull request).
Documented experience
Described in the résumé published with this portfolio; not independently verifiable here.
Self-reported
A metric stated by the candidate (e.g., the PromptOS cost reduction); no raw data published.
Experimental
Early prototype or active development; not a claimed production deployment.
Repository currently unavailable
The published repository link returns 404 (checked 2026-09-30); the description is not independently verifiable.
candidate.json · evidence.json (claim → capability → project → implementation → result → verification) · capabilities.json · agent-card.json · llms.txt · Agent Mode record
■ Additional Coding Activity
LeetCode
Profile
37
Total Solved
(01)
Competitive programming keeps my fundamentals strong. I solve DSA problems to stay sharp.
■ Additional Projects
Also in
the Record
(09–11)
Additional projects
Legitimate experimental work kept in the record. AgentOS is public and independently checkable; Mobile GUI Automation Agent and Pocket Typist are private repositories — anonymous visitors receive a 404, and read access can be provided to recruiters on request, so those two descriptions are the owner's own account.
(09)
Mobile GUI Automation Agent
Private repo · access on requestiPhone MirroringVision + GUI automation
Computer-use automation over iPhone Mirroring combining screen perception, computer vision, and GUI automation. Original implementation: a macOS agent that drove a dating app (Hinge) interface to generate reply suggestions — the Hinge use case is the prototype, the engineering is the mobile GUI automation layer.
(10)
AgentOS
Public repo6 composable layers250+ app integrations
A personal agent operating system with 6 composable layers: Interface, Model Router, Simulation Engine, Memory Graph, Guard Layer, and Execution Plane with 250+ app integrations.
(11)
Pocket Typist
Private repo · access on requestBLE HID keyboard firmwareCloudflare Workers relay
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.
■ Technical Notes
Engineering
Notes
Short technical notes kept on this page as excerpts — not separate published posts.
What I Learned Contributing to a 2,620-Star Project
Merging PR #190 into webcmd — unifying output format parsing across 15+ commands — showed me how an established codebase handles change.
Aug 2026
Treating Prompts as Source Code with PromptOS
A multi-pass compiler for prompts: why I built a 10-pass optimization pipeline and what it taught me about prompt engineering at scale.
Jul 2026
ESP32-CAM: From ML Dataset to Autonomous Robot
Three projects, one chip. How I went from capturing training images to building a color-tracking robot that drives itself.
Jul 2026
Mobile GUI Automation over iPhone Mirroring
Building a computer-use agent that reads a mirrored iPhone screen and drives the GUI with vision + automation — the original prototype targeted a dating app interface.
Jun 2026
DeepOps: Teaching Infrastructure to Think
What happens when you hand an LLM an incident queue? Root cause analysis through causal chains, not just pattern matching — an experimental platform in active development.
May 2026
Email
snehsinha30@gmail.com
GitHub
Sneh30
LinkedIn
linkedin.com/in/sneh30
Résumé
View PDF →
Agent Mode
Structured candidate record →
llms.txt
Machine-readable summary →
Bangalore, India
Open to work
© 2026 Sneh Sinha
v2