Senior AI Engineer • LLM agents, RAG, MCP, evals • TypeScript & Python • Full-stack (React, Node.js) • Buenos Aires, Argentina
✍️ I write about building agents at molchanov.blog
- 🤖 Production coding agent (ROGII): I built and operate an autonomous multi-agent coding system in production. It picks up Jira tasks, writes the fix, passes independent AI review and opens merge requests. I also built its human-in-the-loop guardrails, its Retrieval-Augmented Generation (RAG) layer over project docs and code history, and the evals that measure its quality. Stack: Claude Agent SDK, TypeScript, Docker.
- 🛠 Agent skills for the team: I wrote reusable agent skills that moved QA from manual to agent-driven Playwright e2e testing and sped up onboarding of new engineers.
- 📊 LLM pipelines with evals (independent work): for a B2B parts marketplace I built a tiered pipeline: rules first, then local classifiers, then a hosted LLM with OpenRouter fallback. It runs 20+ AI features with Zod-validated structured outputs and Langfuse tracing, an eval suite over 1,300+ labeled examples, and a read-only Model Context Protocol (MCP) server.
- 🔎 Multi-agent research: I built a Claude Code multi-agent research system for company research. An orchestrator plans the work and checks the results, and 8 subagents research in parallel. One run maps a company's subsidiaries and contractors, then tests hypotheses about it against public sources. An open-source version is in progress.
- 🧩 Before AI, full-stack and UI engineering: I built the React/TypeScript UI for an oil & gas geosteering platform (Canvas 2D well cross-sections, Leaflet to Mapbox migration, ArcGIS integration). At ROKO Labs I cut bundle size 15% and page load time 20%, and built a virtualized data grid for 10,000+ rows at 60fps. Backend: Node.js, NestJS, Fastify, Python/FastAPI, PostgreSQL.
- Body Double — a self-hosted AI tutor built for ADHD brains. It teaches the open course AI Engineering from Scratch one lesson at a time. Early beta.
- qwen-image-apple-silicon-bench — Qwen-Image-2.1 benchmark on Apple Silicon (M1 Pro, 32 GB): GGUF quantizations Q4/Q8/F16 via stable-diffusion.cpp; speed, memory and image quality.
- molchanov.blog — my blog. Astro + React + Feature-Sliced Design on Cloudflare Pages, with a free "Markdown for Agents" layer for AI crawlers and IndexNow on every deploy.
- Multi-agent research agent — orchestrator + 8 parallel subagents, hypothesis checks, SQLite knowledge base. Open-source version coming.
More on molchanov.blog/projects.
Building agents too? I like comparing notes.



