ML Researcher
·
Efficient multimodal models
·
Medical AI
ML researcher working where reasoning has to be
checked rather than taken on faith, currently a
visiting researcher at the
Tübingen AI Center
(Kühne Group), research engineer at
Tanit Healthcare (Paris), and an Engineering
student at École Polytechnique de Tunisie.
I move between research and engineering with a stubborn belief
that the next interesting wave of intelligence will come from
how we shape models: efficient multimodal foundation
models, vision-language compression, post-training recipes,
and medical AI.
03.2026 Joined the Tübingen AI Center as a visiting researcher with the Kühne Group, working on token compression for omni-modal foundation models.
02.2026 Released VulnScout-C, a 693M-parameter MoE transformer for C-code vulnerability detection, outperforming GPT-4o on CASTLE. Under submission at IEEE TDSC.
02.2026 Won 2nd place at GOAT_AI 1.0 (monocular depth estimation) with a 1M-parameter student distilled from Depth Anything V2-Small.
A 693M-parameter Mixture-of-Experts transformer (353M active)
derived from Qwen3, paired with a 33K-sample dual-verified
dataset built through a multi-agent pipeline. Outperforms
GPT-4o, GPT-o3 Mini, and DeepSeek R1 on the CASTLE benchmark
at a fraction of their inference cost.
CSG-Chat: Constructing Causal Structured Graphs from Qualitative Interviews
Bechir Dardouri
A multi-agent neuro-symbolic pipeline for automated causal
graph construction from qualitative interview transcripts,
designed for policy analysis and downstream decision support.
Notes on efficient reasoning, post-training, and the engineering of intelligence.
Drafting · Apr 2026
What We Let Models Be Wrong About
On the design space of verifiable rewards: what we choose to
grade a model on shapes what it optimizes for, and what it
quietly discards along the way.
post-trainingrlphilosophy
Planned · May 2026
Token Compression Without Regret
Notes from the Tübingen AI Center on keeping compressed
representations faithful to fine-grained reasoning, a
practical guide to scaling perceptual context without
scaling sequence length.
Multi-agent system that turns life-story signals into a typed causal DAG.
Counterfactual, confounder, and gap-filler agents refine the graph
through an interactive interview with LLM-judge evaluation.
VEST (Vision-vs-prior Equity Score Test): the first public audit of any
perception-aware VLM RL method. 50 reproducible tests, all probe
parquets committed.
Production-grade RAG for medical research papers built on Mistral-7B-Instruct
(4-bit), BGE embeddings + FAISS, containerized FastAPI service with
Prometheus observability.
Silver-medal Kaggle solution for AIMO Progress Prize 3 using
Apriel-1.6-15B-Thinker with DeepConf-weighted voting and batched
Tool-Integrated Reasoning under a 5-hour budget.
I'm a member of the
Tunisian national sailing team
and treat it as a parallel kind of optimization: read the
conditions, find the lift, sail the cleanest line.
Sail fast, optimize faster.
06
Contact
Open to research collaborations, internships, and conversations
about efficient reasoning, vision-language models, and medical AI.
