Research · Haseeb Raza

Publications

Biologically Inspired Document Embedding in Discrete Spaces

AI research · ELTE AI Department, Prof. László Gulyás

ACM ICPS · IntRob Fair 2026 · oral presentation

I mapped document embeddings onto a discrete semantic grid, then compared the result with PCA, t-SNE, and UMAP.
Boundary-Work by Artificial Intelligence: How Large-Language Models Govern Sensitive Public Discourse Across User Personas

Responsible AI · RC2S2, ELTE · with Ildiko Barna, Renáta Németh, Márton Gosztonyi

Presented · STS NL 2026, University of Twente · Apr 15, 2026

Accepted talk · 17th ESA Conference, Warsaw · Aug 2026 · manuscript in preparation

We tested the same sensitive questions across 15,552 synthetic personas. The model’s answer changed depending on who it thought was asking.
FAAR: Failure-Aware Agentic Retrieval

In preparation

I’m working on typed failure diagnosis for OCR-heavy document QA. The goal is to recover from the actual failure instead of restarting the whole pipeline and hoping.

Things I’m building in public.

Agent Autopsy↗

Python · LangGraph · OpenTelemetry · CLI · MCP

Give it an agent trace and it runs 14 deterministic failure checks, fully offline. You can inspect what broke through the CLI, Streamlit, or MCP.
EngageMind↗

Python · RAG · LoRA · full-stack

My BSc thesis: a document QA app with RAG, LoRA fine-tuning, stateful memory, and checks for hallucinations.
FailSafeRAG↗

Python · OCR · RAG

An OCR + RAG pipeline for document QA. It tries to notice extraction failures early instead of feeding bad text into retrieval and pretending everything is fine.