AI/ML Engineer
I build production ML systems, from fine-tuning open-source LLMs to deploying inference infrastructure. Working on fine-tuning SLMs on Arabic conversational data using SFT, DPO, and LoRA, with training pipelines orchestrated via Airflow and experiments tracked on W&B. Previously built speech processing, TTS, and STT systems for Arabic audio at production scale.
Co-authored the EGY-MER dataset paper (Research Square, 2025).
Languages
ML & Training
SFT · DPO · LoRA · Instruction Tuning · Contrastive Pre-training · Multimodal ML
MLOps & Infrastructure
| Project | Description |
|---|---|
| LLM Cache Proxy | Rust caching proxy for LLM APIs — 4ms cached vs 2.2s live, ~48% cost reduction |
| ML Inference Platform | ONNX serving with canary routing, shadow mode, drift detection, and Grafana observability |
| trainsafe | HuggingFace/TRL callback for behavioral health checks during fine-tuning — catches language drift, output collapse, repetition loops mid-training |
| Speculative Decoding from Scratch | PyTorch impl of Leviathan et al. 2022 with acceptance rate benchmarks across model pairs |
| clamp-cc | TUI for fine-grained Claude Code context compaction — tag turns, generate targeted instructions |
| TEN-VAD Realtime | Real-time speech segmentation with WebSocket event streaming built on TenVAD |
| Orbis | Full-stack AI chatbot with RAG, web search, and streaming chat — Next.js + FastAPI + Groq |