building agents. then figuring out where they break.
cs @ elte · ai engineering intern @ infineon · research assistant @ eth zurich · founder, elte ds&ai club (100+ members)
what i'm building
| project | what it does |
|---|---|
| agent-autopsy | terminal forensics for agent traces — failure loops, root causes, structured fixes |
| AuditLens | fairness auditing for ML datasets — statistical tests + LLM guidance |
| multi-agent-reflection-engine | 5-agent reflection system with eval pipeline and audit snapshots |
| QualModel | turns survey responses and interview transcripts into structured research models |
| obsidian-ai-agent | AI agent for Obsidian — answers grounded in your own notes |
research · papers in prep
| project | what it does | status |
|---|---|---|
| failure-aware-ocr-rag | OCR + RAG pipeline with typed failure handling and recovery routing | active |
| llm-bias-evaluation | persona-based bias eval across LLMs — 15,552 personas @ elte rc2s2 | paper in prep |
| EngageMind | RAG study assistant with chat memory and GPT-2 LoRA — BSc thesis @ elte | almost :) |
research without deployment is theory. deployment without evaluation is guesswork.