📍 Bucharest, EU time | 🤖 Building agent systems that run a real business | 🎯 Open to remote AI Engineer roles
Founder of Ministeru' Creativ, a four-person agency whose back office is an agent fleet I built and operate. BSc Computer Games Design and Development (Worcester, 2022).
- 🧭 icp-agent - agentic lead engagement: enrich a company, assess-loop until confident, ICP score, route, draft the email, log to a Sheet. ReAct with closed-enum actions, every failure mode handled explicitly. Live demo
- 🎫 instantly-support-agent - support-ticket responder: RAG over an indexed help centre, self-graded drafts, a multi-signal confidence gate that decides auto-send vs human review, and the full decision chain on one page per ticket. Live demo (sign-in)
- 📄 cv-tailor - job description in, tailored CV out: one structured LLM call, a deterministic Jinja/WeasyPrint renderer, an honesty guard that refuses invented experience, browser-faithful host allow-listing before any portal submit, leak-canary tests. 539 tests.
- 🚿 faraapacalda.ro - how many days a year Bucharest goes without hot water. A 15-minute git-scraper (termo-data) feeds a 2,350-page static site (termo-site) with real users and a feedback loop.
- 🎧 ColdCoach - live coaching for cold calls, native macOS: on-device WhisperKit transcription, coaching cards within ~2 s of an objection, a playbook that learns from logged outcomes, bring your own key. Homebrew tap, CalVer releases.
- 🧠 wiki-substrate - the memory substrate my agents run on: a typed markdown schema, ingest and lint tooling, and MCP wiring over Tobi Lutke's qmd.
- 🧭 icp-agent - enrich, assess, score, route, email, log; an agent that knows when it has enough
- 🎫 instantly-support-agent - RAG plus a confidence gate that knows when to hand off to a human
- 📄 cv-tailor - structured output, deterministic rendering, an honesty guard
- 🧠 wiki-substrate - schema and MCP wiring for a markdown vault used as agent memory
- 🗄️ termo-data - git-scraper: CMTEB outage pages snapshotted every 15 minutes since June 2026, committed only on change
- 🚿 termo-site - pure-SSG Next.js site over that archive, street and heating-point level, 122 tests plus Playwright e2e
- 🎧 ColdCoach - Swift, WhisperKit, floating overlay, Keychain-stored keys
- 🍺 homebrew-coldcoach -
brew install --cask tiXor-code/coldcoach/coldcoach
- 🚀 eve-arbitrage - inter-region hauling arbitrage in EVE Online, ranked by exact profit per trip after walking the real ESI order books. Live
- 📐 room-planner - drag-and-drop room planner on react-konva: rooms, furniture, doors, snapping, undo/redo, JSON import/export. Live
- 🏚️ Puzzle-Haunt - final-year thriller puzzle game set in the Worcester Royal Infirmary, on a hand-rolled MonoGame engine (C#)
- 🏰 The-City - tower defence in Unity, built for fun
Most of what I build runs my own company and stays private:
- AIOS - the agent fleet behind Ministeru' Creativ: a health sentinel with heartbeats, a 298-rule secure development reference enforced by four scanners that file findings as issues, an invoice auto-filer, a Telegram job spine running headless Claude workers, and a phone-first decision cockpit. ~970 tests.
- Trading research - automated micro-futures and prediction-market strategies with a settlement-truth ledger and a fail-closed risk layer.
- Client work - sites, booking systems and lead pipelines for agency clients, through Ministeru' Creativ.
- Spec, then plan, then review-gated implementation by subagents. Every build is test-gated and the test count is part of the report.
- Headless agents on launchd with heartbeats, a health sentinel and quiet hours. Nothing runs armed without a kill switch.
- A security gate on every change: a written rules reference is checked before a task is called done, and every finding carries an ID and a severity.
- A 700-page markdown wiki is the operating memory for every session. The agents read it before they act.
