Houssem Eddine Ghribi

AI Researcher & Engineer | Explainable AI · LLMs & RAG · Multimodal & Agentic Systems

Final-year AI engineering student at ESPRIT University (Master's-level, graduating 2027) with 2+ years of industry AI work for companies in Germany and France.

I am driven by one question: does this AI system actually work, and why? I build real systems (LLM and RAG platforms, vision models, multi-agent pipelines) and then open them up: causal probing to see what happens inside a model, controlled ablations to see which design choices matter, and pre-specified evaluation to make sure a result is real and not noise.

The domain changes, the method does not. Whether it is a vision-language model reading an X-ray, a RAG system answering from contracts, an LLM writing clinical reports, or a vision model inspecting a production line, I ask the same three questions: what is the model actually using, which choices really matter, and would the result survive a fresh test? That thread connects my work across explainable AI, LLMs, RAG, NLP, computer vision, and medical imaging.

🎯 Seeking: a PhD position starting Fall 2027 in AI (XAI, LLMs, RAG, NLP, multimodal, or applied AI). Also open to a 6-month research internship from January 2027.

• Explainable & interpretable AI: what really happens inside LLMs and VLMs, using causal methods (attention knockout, information flow, representation analysis).
• LLM reliability, RAG & NLP: hallucination, faithfulness, grounding, and evaluation methodology (golden datasets, LLM-as-judge, ablations).
• Multimodal & agentic systems: models and multi-agent architectures that combine vision, language, and structured reasoning.
• Applied AI with real stakes: healthcare, industry, and regulated domains where errors cost something.

All papers and code are open. Full list on ORCID. More in progress.

Year	Title	Topic	Links
2026	Medical Fine-Tuning Changes Answer Commitment and Information Flow in a Vision-Language Model: A Split-Sample Study on VQA-RAD Houssem Eddine (single author)	XAI · VLMs · Interpretability

• Inside the model: causal interpretability for LLMs and VLMs. Which layers and components actually drive an answer?
• Trustworthy LLMs & RAG: detecting when a model answers from prior knowledge instead of the evidence it was given.
• Better evaluation: protocols that separate real improvements from noise, for any AI system.

🔬 Causal interpretability of vision-language models (paper above) Turned attention-knockout probing into a before/after measurement instrument with built-in replication. Architecture-agnostic, works on any HuggingFace VLM, no retraining.

📚 Evaluation-driven RAG reliability | Alight Consulting, Germany Question: can a lightweight LLM match a frontier model on grounded question answering, and how do you prove it? Built a golden-dataset harness (factual, analytical, refusal, access-control) scored by an LLM-as-judge rubric, with controlled ablations (HyDE on/off, compression sweeps) over 749 graded runs across 10 configurations. Finding: a small model reached frontier-level quality (96.8% pass) at ~7x lower cost.

🫀 Clinically validated 3D segmentation | Alight Consulting, Germany 3D cardiac MRI segmentation (MONAI) validated by a blinded study with 2 senior cardiologists across 40 cases: Dice 0.88 to 0.91, clinically acceptable in 95%+ of cases.

Azure AI Search · Azure OpenAI · Document Intelligence · FastAPI · Next.js · LLMLingua

• Architected a production RAG platform: 4-stage ingestion, hybrid BM25 + HNSW retrieval with Reciprocal Rank Fusion, adaptive HyDE routing (faithfulness 4.96/5).
• Index-level access control validated at 100% compliance, zero leakage across 31 cases and 7 group configurations.
• Evaluation harness (749 runs, 10 configs): chose a model matching gpt-4.1 quality at ~7x lower cost, cost-modeled to ~$496/mo at 500 users.
• Full-stack app with multi-turn conversation and a PDF viewer that opens on the exact cited page.

YOLOv8 · OpenCV · LangChain · LangGraph · FastAPI · Playwright · A2A · MCP

• Computer vision: mobile CV pipeline on a wiring harness line cut return rate from 6.8% to 0.2% in 30 days.
• Agentic GTM (freelance): delivered a one-click lead intelligence app; a two-hop channel graph turned 400+ targets into ~10 high-leverage contacts.

• Built GreenOPS AI, a human-in-the-loop energy optimization system for a pharmaceutical plant. 2nd place, Insight for Impact.

MONAI · PyTorch · SFT + DPO · ONNX

• 3D segmentation pipeline: Dice 0.88 to 0.91, 18% over baseline, clinically validated.
• Fine-tuned an LLM (SFT + DPO) on 200+ expert-reviewed reports: reporting time -30%, hallucination rate -35%.

• 92.4% IoU on multimodal satellite imagery, 70% less manual labelling, 31% better generalisation to unseen terrain.
• CrewAI multi-agent reporting: ~10 hours of weekly work down to ~20 minutes.

• Forecasting error (MAPE) 48% → 12%, unsold inventory -42%, VLM-based product analysis +20% conversion.

• Promoted from intern to full-time in one month after a model cut food waste by 55%.

XAI & Evaluation: attention knockout · information flow · GradCAM · CKA · ablation design · LLM-as-judge · pre-specified testing LLMs, RAG & NLP: LangChain · LangGraph · CrewAI · hybrid retrieval · HyDE · LLMLingua · SFT + DPO · A2A · MCP Vision & Imaging: YOLOv8 · U-Net · MONAI · SimpleITK · DICOM · 3D Slicer

Diplôme d'Ingénieur, Computer Science (AI Specialisation) | ESPRIT University | 2022 to 2027 Master's-level degree (EQF Level 7). GPA: 15/20 (French scale), approximately 3.7 / 4.0.

Founder & Tech Manager | DeepFlow | Jun 2023 to Present AI education initiative: 20+ workshops (deep learning, LLMs, medical imaging) for 200+ engineering students; mentored 24 hackathon competitors.