Malek
Senoussi

KAN RL Environment — LLM agent for scientific law discovery

An LLM agent trained via a reinforcement learning loop to autonomously fit Kolmogorov-Arnold Networks for symbolic regression. Structured feedback on generalization, OOD extrapolation, and expression parsimony — with no gradient flowing through the LLM.

→ Evaluation loop architecture for autonomous agent training. A building block for research systems where LLMs train specialized models without human supervision.

Board Game Generator — Multi-agent pipeline

A LangGraph pipeline orchestrating 5 specialized Claude Sonnet agents that generate fully playable browser games from a single text prompt. Each agent handles a distinct stage — game design, code generation, SVG assets, QA validation, documentation — with an automatic retry loop when the Tester detects blocking bugs.

→ 3 games generated end-to-end. €10 in API costs. Zero manual fixes on pipeline v2 vs 5 on v1 — a measurable robustness improvement across iterations.

→ Multi-agent orchestration with a stateful LangGraph graph, structured output, retry loops, JS syntax validation, and full CI/CD. Games are live and playable.

LLM Diagnostic Framework

Systematic framework for diagnosing LLM failures and testing optimization strategies in production. Case study on medical entity extraction; modular and extensible to other use cases.

→ Reduced inference cost by 10× while maintaining accuracy (91% vs 92%) on medical entity extraction.

Pick the right model for the right complexity and cost.

High-dimensional classification pipeline

Deep learning models for biological data with 20K+ features. Custom PyTorch architectures with automated feature selection and dimensionality reduction, achieving state-of-the-art performance on complex scRNA-seq classification.

→ State-of-the-art performance on real scRNA-seq datasets. Benchmarked against existing methods in peer-reviewed work (see Publications).