I build the platforms that run AI agents in production.

A premium Windows tray utility that turns webcam activity into a click-through ambient screen glow. Designed for privacy visibility without opening or reading your camera feeds.

Problem

Webcam active lights are tiny, single-monitor, and outside your line of sight. It is easy to miss background applications leaving your camera active during games, streams, or workspace multitasking.

Approach

Wrote a lightweight C# tray app that polls Windows ConsentStore registry keys and hooks WMI process triggers. Created a translucent, click-through WPF viewport border that displays responsive ambient alerts across all active screens.

Outcome

Zero-telemetry, zero-pixel-read background utility sold via Razorpay license validations and lifetime seat purchases. Active releases hosted at fits.ninja. Registry monitorsWMI FallbackMulti-display routing₹200.00 / seat

• C#
• .NET 8
• WPF
• Windows API
• Registry Hooks
• WMI Queries
• Inno Setup 6

Interactive Preview ↗ Buy on fits.ninja ↗

The control plane that sits above AI agents doing production work — provisioning them, streaming their output live, gating irreversible actions behind human approval, and auditing everything they do.

Problem

AI agents in production are blind spots: nobody can see what an agent is doing right now, stop it before an irreversible action, or audit what it did last week. Tracing one production API error to root cause took an engineer ~3 hours.

Approach

Built the platform solo across 45+ backend service modules: Claude CLI subprocesses streamed over Socket.IO, Board/Task/Approval workflows for human sign-off, and an RCA pipeline that pulls SigNoz error logs, reads Bitbucket source, and reports every conclusion with an evidence chain down to file, line, owner, and confidence score.

Outcome

3-hour manual traces now run as 10-minute automated RCAs — in production for a fintech engineering team. ~18K LOC175+ tests50+ REST endpoints40+ Prisma models25+ WebSocket events

• Node.js
• Express
• TypeScript
• PostgreSQL
• Prisma
• Socket.IO
• React
• Tailwind
• Claude CLI
• MCP

A canvas where LLM and media pipelines are wired as a visual graph and executed by background workers — multimodal automation without brittle scripts.

Problem

Multimodal pipelines — prompt, crop an image, extract video frames, run a vision model — usually live in one-off scripts that nobody else can run, inspect, or reuse.

Approach

React Flow canvas with 6 typed node types (text, media upload, Gemini vision, image crop, video frame-extract), typed ports with cycle prevention, undo/redo, JSON import/export and debounced autosave — executed topologically by Trigger.dev workers running FFmpeg and Gemini, behind Clerk auth and Zod validation.

Outcome

Per-user run history in PostgreSQL/Prisma — single-node, partial, and full-graph runs each auditable with node-level inputs, outputs, and timings. 6 node typesDAG validationfull run audit

• Next.js 16
• React Flow
• TypeScript
• Trigger.dev
• Gemini
• FFmpeg
• Transloadit
• Prisma
• PostgreSQL

Live demo ↗ GitHub ↗

A citation-grounded assistant that answers fixed-deposit questions by voice or text in 15 Indian languages — every answer traceable to its source paragraph.

Problem

Financial guidance is worthless when it hallucinates — and English-only interfaces lock out most Indian users entirely.

Approach

Hybrid retrieval: BGE-M3 dense vectors in FAISS fused with BM25 via Reciprocal Rank Fusion, MMR diversification, and cross-encoder reranking. Intent routing, compliance redaction, and deterministic financial math keep numbers computed — never generated. Web Speech STT/TTS with language-specific UI state, source drawers, and evidence highlighting.

Outcome

15 languages spoken and written, retrieval telemetry on every query, and a 35-case multilingual evaluation suite tracking answer quality. 15 languages35-case eval suitecited answers

• Next.js 15
• React 19
• TypeScript
• Python
• FastAPI
• FAISS
• BGE-M3
• BM25
• Ollama
• PostgreSQL

GitHub ↗