What I work on: how a system arrives at an answer — and how faithful that process is to the evidence behind it. I study it across human gaze (scanpaths), model reasoning traces, and retrieval-grounded systems, and I build the measurement harness that checks the answer rather than trusting it.
I'm an AI/ML researcher in Paris — AI Researcher at F.initiatives and AIOps Architect at NeoPhi.
🔭 Currently: foveated multimodal models and the faithfulness of machine vs. human gaze · retrieval and systematic-review agents at NeoPhi • 🌱 Interests: visual attention · trustworthy & interpretable AI · LLM post-training • 💬 Ask me about scanpath modeling, RAG faithfulness, and model merging
- SP_Gen — domain-adaptive generative model for scanpath prediction on paintings · CBMI 2022
- SSLArtScanpath — self-supervised (Barlow Twins) scanpath prediction · CVPR Workshops 2022
- FDISS — Foveal Disc IoU Scanpath Score, a biologically grounded scanpath metric ·
pip install fdiss - scanpath_nlp_metrics — compare scanpaths via VLM descriptions + the classical spatial & MultiMatch suite ·
pip install scanpath_nlp_metrics - AVAtt — visual-attention dataset for paintings (scanpaths + saliency)
- ThinkProb — profiles LLM reasoning by turning Chain-of-Thought into Thought Graphs (non-generative) and scoring a 22-dimensional cognitive profile
- HeReFaNMi — Health-Related Fake News Mitigation: tackles online health-news misinformation by retrieving evidence and triaging claims as trustworthy / doubtful / fake, with source validation · EU NGI Search–funded
- Enlighten-AI — a mental-health & life-coaching assistant grounded in Dr. K's (HealthyGamer.gg) video transcripts: citation-grounded RAG answering with timestamped sources, safety guardrails, and a red-team report
- llm-tuning-lab — 11-skill toolkit for post-training LLMs as controlled experiments (SFT, LoRA/QLoRA, CPT, preference optimization, merging) with a significance backbone
- merge-corrected SFT distillation — reasoning distillation into Qwen3.5-0.8B with per-epoch, benchmark-gated merge correction · model on 🤗 Hub
- APEX — agentic portrait editing where an MLLM plans & judges each edit and an ArcFace identity gate must agree before an edit is accepted
- M. Tliba, M. A. Kerkouri, A. Chetouani, A. Bruno. Self-Supervised Scanpath Prediction Framework for Painting Images. CVPR Workshops, 2022. — paper
- M. A. Kerkouri, M. Tliba, A. Chetouani, A. Bruno. A Domain-Adaptive Deep Learning Solution for Scanpath Prediction of Paintings. CBMI, 2022. — paper
Full list on Google Scholar and ResearchGate.