- 🤖 I build production-grade LLM systems — RAG fusion pipelines, multi-agent DAGs (LangGraph), and local-first QLoRA fine-tuning with auto device detection (CUDA / MPS / CPU)
- 🏥 I work on clinician-supervised clinical AI — multimodal triage agents on Gemini Live API with structured JSON outputs across 9 languages
- 📊 I'm evaluation-obsessed — every system I ship has a
/benchmarkendpoint, JSON-parse rates, MAE per dimension, or tier-accuracy reports - 🛡️ I architect for responsible AI — decoupling bias detection from scoring, separating observation from influence, designing for auditability
- 🛣️ My foundations are in classical CV — OpenCV + Hough transforms, real-time stream processing with confidence scoring
- 📍 Based in Darmstadt 🇩🇪 — currently exploring research collaborations in healthcare AI and applied ML
LLM & Agents
RAG & Data
ML / CV / NLP
Backend & Web
Languages & Infra
|
LLM-powered resume screener with RAG, multi-agent DAG, bias decoupling, and local QLoRA fine-tuning. 5-dimension weighted scoring · RAG Fusion (Qdrant + BGE-large) with Reciprocal Rank Fusion · LangGraph DAG running RAG ∥ bias in parallel · deterministic composite math · QLoRA pipeline auto-detecting CUDA/MPS/CPU.
|
Clinician-supervised, non-diagnostic multimodal AI for overloaded clinics. Voice + docs + text, 9 languages, 6 modules. Built on Gemini 2.0 + Live API · structured JSON triage with severity-coded UI · role-aware prompts (Clinician / Patient / Caregiver) · WebSocket live voice via LiveClient · graceful Markdown fallback when JSON parsing fails.
|
|
TextBlob lexicon vs Gemini 2.0 Flash — side-by-side with a benchmarking endpoint. Two engines on the same input · agreement flag + polarity delta ·
|
Real-time CV pipeline with structured per-frame stream logging and confidence scoring. OpenCV + Canny + Hough transform · per-frame JSON records (status, confidence, fps) · DETECTED / PARTIAL / NO_LANES tiers · summary metrics on exit · annotated video output. Treats CV like a streaming data system, not a script.
|
🔬 More projects — research & experiments
- Beyond the Siren — Evaluation framework for equitable emergency care in low- and middle-income countries
- MLJAK2-Biotech — ML applied to biotech / JAK2 research workflows
- Company Culture Analysis — NLP-driven culture analysis from employee reviews using sentiment scoring, tokenization, and word-frequency visualization
- Maxwell's Rule in AR — WebXR experience visualizing electromagnetic theory ⭐
|
|
|
Open to collaborations and roles in LLM systems engineering, applied ML research, and responsible healthcare AI.
If you're shipping evaluation-first ML or building agentic systems with care for safety and bias — let's talk.