Production AI features, built to a spec and handed back tested. Agents, LLM integrations, internal tools and the infrastructure they run on.
I work as the build side for studios and product teams: you own the client and the scope, I return a working, tested, deployed piece. Written async communication, fixed scope per task.
Site · Live CRM demo · Telegram channel
| Project | What it proves | Stack |
|---|---|---|
| crm-pro | A full B2B sales CRM with six role workspaces, an AI assistant, voice-to-document and an ERP integration layer. Try the demo | Next.js 16, React 19, Prisma, PostgreSQL, clean architecture |
| citewise | Document Q&A where every sentence carries a citation that is checked against the source in code, with an eval suite that reports its own misses. Runs with no API key. Try the demo | TypeScript, Express, BM25 |
| docfields | Invoice extraction where every number is validated in code (line items, subtotal, tax, total) and reviewed next to the text it came from. Try the demo | TypeScript, Express, PDF parsing |
| vault-research-agent | A multi-agent research pipeline with an eval suite: planner, router, parallel MCP workers, synthesizer and an LLM judge that sends weak answers back. Runs from a clean clone with no API key. | TypeScript, Claude, MCP |
| benchmark-cc | Measurement instead of opinion: an A/B harness that isolates what each layer of a Claude Code setup contributes in tokens, wall time and judged quality. | Node.js, Claude Code |
| btw-agents-pack | 37 subagents and 17 skills extracted from daily production use, each with an explicit tool allowlist. MIT. | Claude Code |
| agent-dashboard | Real-time observability for running agents, drawn as a pixel-art office. Live session tracking over WebSocket with a polling fallback. | Node.js, Canvas 2D, WebSocket |
| Nox (showcase, code is private) | A personal AI assistant in daily use: Telegram bot, Android app with over-the-air updates, scheduled jobs, alerting, and a self-hosted server it monitors itself. | TypeScript, grammY, SQLite, Capacitor, Claude Agent SDK |
- Spec first. Requirements and out-of-scope are written down before code.
- Tests and a real run. A change ships after the test suite passes and the path was exercised in a browser or against the live system.
- AI behind a port. Model providers and external systems sit behind interfaces, so swapping a provider or a stub for the real thing is a one-file change.
- Self-hosted when it pays. Docker, Cloudflare tunnels and a monitored home server run the demos linked here.
AI Claude API and Agent SDK, MCP, tool use, evals with LLM judges, OpenRouter, Gemini Backend TypeScript, Node.js, Prisma, PostgreSQL, SQLite, grammY Frontend Next.js, React, Tailwind, Phaser, react-three-fiber Infra Docker, Caddy, Cloudflare, GitHub Actions, Vercel
Subcontract builds for studios and agencies, part-time contract work, AI feature integration into an existing product.


