I build backend-heavy AI systems which include document intelligence pipelines, multi-agent orchestrators, hybrid rule+model architectures. My focus is on the engineering layer: explainable outputs, structured schemas, and pipelines where every decision is traceable.
Interned at SFI Data (Amsterdam). Won Neurax Hackathon — 1st place.
DocStruct — Hybrid PDF layout analysis pipeline. Geometry proposals and DocLayNet DETR detections generated independently, fused via IoU, classified with per-source confidence breakdowns. mAP@0.50 0.844 on DocLayNet v1.2. Full evaluation harness + pytest suite.
Neurax-CoFounders 🥇 — Multi-agent workforce orchestrator. Reads employee resumes, decomposes a PRD into atomic tasks, skill-matches them, and executes via spawned CrewAI subagents. Handles skill gaps by emitting structured hiring alerts. 1st place.
SFI-Data-Project — Production pipeline built at SFI Data. Dual-track LLM extraction from 50–200 page corporate sustainability PDFs: vector retrieval for narrative fields, full-context pass for tables. Output: structured Excel for financial analysts.
ClearPath · SeizureSense · clinical_health
Python FastAPI Flask PyMuPDF HuggingFace Transformers CrewAI Pydantic SQLAlchemy React hls4ml
