I build the full lifecycle of AI systems — from data pipelines to model training, evaluation, and deployment that actually runs in production. Lately focused on ML, GenAI and AI Safety.
data→train→evaluate→ship→monitor
Six years building production systems, the last two focused on artificial intelligence, Generative AI, and AI safety. A strong software engineering foundation (C/C++, École 42) underpins everything — I'm comfortable across the whole pipeline, from low-level systems to model design.
A mix of production systems, research, and end-to-end ML projects. Code on GitHub.
End-to-end GenAI safety pipeline: data generation → LoRA/QLoRA finetuning → automated evaluation across a 20+ category harm taxonomy. Raised Attack Success Rate from 38.8% → 65.9% on held-out categories. Multi-agent harness with an LLM-as-Judge, plus interpretability research to detect multi-turn jailbreak attacks.
Benchmarked CLIP vs BiomedCLIP for pneumonia detection on chest X-rays. BiomedCLIP + LightGBM matched supervised baselines with only ~140 labeled images. Added an LLM-as-Judge quality gate to flag poorly-captured scans.
Two-phase pipeline: facial-attribute extraction → trait voting → diffusion generation. Benchmarked CNN, ResNet-18 and ViT-B/16 (0.883 acc); generated pets with Stable Diffusion + IP-Adapter, accelerated via LCM-LoRA. Served end-to-end.
AWS-native data platform with medallion architecture (Bronze/Silver/Gold): scraping → Glue ELT → schema-validated S3/RDS → H2O AutoML on SageMaker → served via API Gateway + Lambda.
Entropy-based unsupervised anomaly detection on network traffic (KDD Cup 1999), distinguishing normal patterns from attack types. Benchmarked against supervised baselines with full statistical evaluation.
Image-based sentiment analysis using Vision Transformers combined with LightGBM and Optuna hyperparameter optimization for emotion understanding.
A basic command-line interpreter written from scratch in C: parsing, process creation (fork/exec), PATH resolution, built-in commands, and environment handling — a ground-up implementation of how a shell works.
École 42 project: reads a height-map file and renders it as a 3D wireframe using isometric projection. Built in C with MiniLibX — line drawing (Bresenham), coordinate transforms, and memory-managed parsing from scratch.
École 42 project: a from-scratch implementation of the C standard printf function, handling variadic arguments, format specifiers, flags, and edge cases — a deep dive into low-level string formatting and memory.
A developer who grew into machine learning by shipping it