Bogdan Truta

About

AI Engineer at ERNI and co-founder of AlphaProve, based in Cluj-Napoca. I build production AI systems, quantitative infrastructure, and autonomous-agent tooling. I took 1st place (Application Development) at PoliHack 2025 and 2nd place (AI App Development) at PoliHack 2026.

Projects

• AlphaProve: a deployed multi-tenant quantitative backtesting SaaS I co-founded. I designed the AI loop that turns plain-English strategy descriptions into backtest-ready code, the engine that simulates an exchange at tick precision with L2 order-book fills, and the self-repairing market-data store behind it.
• eGata: an AI agent that walks Romanian citizens through municipal procedures over voice, text, and telephone. Forms are rendered to PDF through LaTeX, every milestone lands in a hash-chained, Ed25519-signed audit ledger the citizen can verify, and which tools the agent may call is decided in code from the current state. Built at Cluj Hackathon 2026 (landing page).
• seeks: a Claude Code plugin that keeps an agent working toward one goal until checks you define pass. A Stop hook runs those checks itself instead of trusting the model, and edits to existing tests send the result to a human. Node.js with no dependencies.
• GOSHA Jobs: a deployed job-matching platform that ranks openings against a candidate's CV with sentence-transformer embeddings and Rocchio relevance feedback (source).
• Prop Firm Trading Calendar: a live calendar feed of maintenance windows and market closures for FTMO, Topstep, Blueberry Funded and E8 Markets. Each announcement is extracted several times by an LLM and an event is kept only if most runs agree; quotes the model gives are checked against the page and mismatches are flagged. The RFC 5545 feed writer is written from scratch (source).
• LearnLoop: a prompting coach that scores prompts across Claude and Copilot. Hono/PostgreSQL API, custom MCP server, Chrome and VS Code extensions, Next.js dashboard. 2nd place (AI) at PoliHack 2026.
• HistoricalCryptoDataDB: a market-data warehouse on TimescaleDB with automated gap repair, an escalating restart-ladder watchdog, a written disaster-recovery runbook, and a nightly scan that re-downloads days with missing data from the exchange archives.
• TradingStats: a quantitative research framework. Out-of-sample hypotheses are committed to git before outcomes are joined, an automated gate rejects model payloads that leak price, dates or asset identity, and refuted results are recorded alongside confirmed ones.
• CS2 Weekly Item Journal: a computer-vision pipeline that reads Counter-Strike 2 weekly item drops from screenshots with a custom-trained YOLO model and OCR, prices them from a daily Steam Market crawl, and recommends the two worth claiming.

Work experience is on the Experience page. Download my CV (PDF), or find me on GitHub and LinkedIn.