Agentic engineering inside the corporate perimeter — AI systems from prototype to production.
I build AI systems that survive enterprise constraints: on-prem models, audit logs, human approval, rollback. Agents, RAG, evals — and the Docker infrastructure they run on.
| Step | What it means |
|---|---|
| 1. Problem → a number | Not "adopt AI", but "cut approval from 3 days to 20 minutes". |
| 2. Sandbox first | Prototype on synthetic data inside an isolated perimeter. |
| 3. Evals before scale | Quality metrics and guardrails go in before rollout, not after an incident. |
| 4. Ship with an owner | Access rights, audit logs, a rollback path and a named owner of the process. |
- VisionQwen — real-time AI assistant for Meta Ray-Ban glasses: Qwen VL vision plus an agentic tool backend. MIT, fork of VisionClaw.
- gigaam-api — stereo call recording in, speaker-separated transcript out: GigaAM ASR, Silero VAD and per-utterance emotion scores behind FastAPI.
- playwright-server — headless Chromium in a container, driven over Playwright Server WebSocket, with a healthcheck that actually opens a page.
- Internal systems (private) — AI content checking, Telegram digests, deck generation from LLM plans, property-data lookups. Python, FastAPI, Node.js, Docker Compose, CI on GitHub Actions.
Python · FastAPI · Node.js · Docker Compose · PostgreSQL/pgvector · Ollama ·
OpenAI-compatible APIs · Playwright · GitHub Actions · Linux (ufw, fail2ban, sshd)
- Telegram: @tr1one