Debajyoti Saikia

Human-like desktop automation. Screenshots the screen, asks an LLM what to do, and drives the real mouse and keyboard — then records tasks as a semantic action graph and replays them deterministically with drift detection. Go standard library only, with a default-deny whitelist, audit log, and MCP server.

Autonomous web agent + AI workspace in one Chrome extension. Plans and executes tasks across sites with local-first privacy and sandboxed code.

A work digital twin. Captures how you think, execute, and what you know from your docs, notes, code, and voice — then lets your own AI agents recall it over MCP and act as you. Method and pattern, never a copy; owner-controlled and encrypted.

Speak a thought or photograph one — a whiteboard, a receipt, a page — and get back a titled, tagged, dated Markdown note that files itself. AI runs exactly once, at capture; reading and searching hit SQLite FTS5 instead of a model, so they stay database-fast. Every note exports as plain Markdown you own.

AI-native language learning platform. Dynamically generated lessons, voice tutoring, spaced repetition across 13+ languages.

AI-generated experimental radio. A control system plans, sources, and bounds each broadcast while Gemini Live supplies native-audio hosts on an independent 48 kHz master clock — original talk, sourced updates, and speculative drama, published to YouTube and rights-aware by design.

Browser-based green-screen keyer and channel editor, built on its own channel-separated container (Prism T1) and a from-scratch lossless codec (PxC1). Splits any clip into R, G, B and alpha tracks, cuts the background with a multi-sample chroma keyer, composites over new footage, and exports with audio, every frame and true alpha preserved — and it's fully drivable by an AI agent over MCP.

VS Code extension that turns your GitHub Copilot Chat history into reusable specs. Reads every request, convention and correction you've typed, groups them by area, and distills each into a working prompt — plus instruction files, a project decision log, and a 21-chart dashboard of how you actually work. History stays local; classification runs on your own Copilot subscription.

An AI agent that paints your photograph one brushstroke at a time — not a filter or style transfer. It picks a palette, tones a ground, blocks in the masses, then looks at its own canvas and decides the next pass. A WebGPU engine simulates real bristles, wet paint that muddies where strokes cross, lit impasto and a hand that isn't quite steady. Runs entirely in your browser on your own model key.

Autonomous market-research agent. Scans live demand and competition to find a low-competition niche for your skills, then hands you a build + go-to-market plan with cited sources.

AI stock analysis agent. Real-time market data with AI-generated insights and portfolio tracking.

Portfolio platform for humans and AI agents. Markdown pages, 15 templates, REST API, and semantic discovery.

Encrypted, ephemeral peer-to-peer file transfer. Files move browser-to-browser over WebRTC with end-to-end encryption — one-time links, no accounts, nothing stored.

Windows screen capture tool for region recording, smart frame capture, AI voiceover, and auto-generated captions for polished demos.

Contactless vital signs from a webcam, computed entirely in the browser. Remote photoplethysmography pulls a pulse waveform from tiny colour changes in facial skin — heart rate via POS projection and FFT peak-picking, plus HRV, drowsiness (PERCLOS), stress and liveness. No server, no upload, no account; every metric is published with its method and its limits. Not a medical device.