AI Engineer @ ERNI · Co-Founder of AlphaProve
I'm Bogdan-Alexandru Truța, an AI engineer in Cluj-Napoca, Romania. At ERNI, I work on a medtech pilot that finds traceable evidence for test steps in approved documents. I also co-founded AlphaProve, where I build the AI strategy workflow and backtesting infrastructure.
Claude Opus 5.5 played Baldur's Gate 3 from a new game to the epilogue with nobody at the controls: 38.9 hours of play and $712 in model costs. The agent sees only what a human player could see.
My contribution: The Lua mod inside the game, the MCP server the agent plays through, a supervisor that launches the game and recovers from crashes, and a 28-scenario benchmark with graders and statistics.
Turn a trading strategy description into code and test it against historical market data.
My contribution: AI strategy workflow, simulation engine, and market-data infrastructure.
Matches job postings to a CV and adjusts recommendations using user feedback. Deployed and used by 20+ students.
My contribution: Embedding-based ranking, Rocchio relevance feedback, and ingestion from eight job boards.
Runs Claude Code toward one goal; a Stop hook re-runs your checks itself before it accepts done.
My contribution: Agent loops, a hook-enforced done gate, test-tamper checks, and file-edit guardrails in a dependency-free Node.js plugin.
A hackathon prototype for Romanian municipal procedures, combining retrieval, voice and text interaction, and tool permissions enforced in code.
An AI prompting coach with Gemini-based scoring, a shared team knowledge base, and integrations across browsers, editors, and MCP. 2nd place (AI) at PoliHack 2026.
Maintains multi-year crypto market data in TimescaleDB with automated gap repair, recovery monitoring, and a disaster-recovery runbook.
Tests trading hypotheses that are pre-registered in git before any outcome is joined, gates every LLM payload against price, date and asset leakage, and records refuted results alongside confirmed ones.
Publishes maintenance windows and market closures for four prop firms as a live calendar feed, using LLM extraction that is majority-voted across repeated runs.
Reads weekly Counter-Strike 2 item drops from screenshots with a custom-trained YOLO model and OCR, then prices them from a daily Steam Market crawl.