Research

Accepted

• › Biologically Inspired Document Embedding in Discrete Spaces

  AI research · ELTE AI Department, Prof. László Gulyás

  ACM ICPS · IntRob Fair 2026 · oral presentation

  Discrete semantic grid mapping of document embeddings, benchmarked against PCA, t-SNE, and UMAP.

• › Divergent Framings: Persona-Dependent LLM Responses on Antisemitism-Adjacent Issues

  Responsible AI · RC2S2, ELTE

  ESA RN31 · 17th ESA Conference, Warsaw (25-28 Aug 2026)

  STS NL 2026, University of Twente (April 2026)

  Paper in preparation

  Post-deployment: persona-driven behavioral bias evaluation across LLMs (15,552 personas).

In Progress

• › FAAR: Failure-Aware Agentic Retrieval

  Paper in preparation

  Typed failure diagnosis and targeted recovery for OCR-heavy document QA pipelines.

Projects

Some of the things I did :)

• › Agent Autopsy↗

  Python · LangGraph · OpenTelemetry · CLI · MCP

  Local-first debugger for LLM agent pipelines. Ingests traces, auto-detects 10+ failure classes, surfaces fixes via CLI, Streamlit, and MCP.

• › EngageMind↗

  Python · RAG · LoRA · full-stack

  Full-stack RAG platform with LoRA fine-tuning, stateful memory, and hallucination mitigation. BSc thesis.

• › AuditLens↗

  Python · FastAPI · Streamlit · fairness auditing

  Dataset-level fairness checks for tabular ML with statistical tests and optional LLM interpretation.

• › FailSafeRAG↗

  Python · OCR · RAG

  OCR + RAG pipeline that handles extraction failures gracefully for document QA, related to ongoing FAAR work.

• › Multi-Agent Reflection Engine (Forkpoint)↗

  Python · multi-agent · evaluation

  Five-agent reflection system with eval pipeline and audit snapshots.