Applied machine learning engineer focused on training, debugging, and deploying deep learning systems, with a particular interest in LLMs and ML infrastructure.
- 🔥 Trained language models from scratch under tight compute constraints
- 🩺 Worked through real training failure modes: plateaus, instability, wasted runs
- 🛠️ Build tooling that makes model training more reliable and efficient
- ⚙️ Comfortable at the systems layer: distributed training, kernel optimization, on-prem GPU infra
- 🌍 Ship multilingual speech and language systems for code-switched, real-world audio
- 🤝 Active open-source contributor on Hugging Face and GitHub
class Salah:
def __init__(self):
self.role = "Applied ML Engineer"
self.focus = ["LLM pretraining", "ASR & diarization", "ML infrastructure"]
self.stack = ["PyTorch", "FSDP", "CUDA/Triton", "FastAPI", "Docker"]
self.happiest = "when loss curves go down and GPUs stay at 100%"
def current_mission(self):
return "Making model training boring: predictable, observable, efficient"| Area | What that looks like |
|---|---|
| 🧬 LLM Training | From-scratch pretraining, continued pretraining, SFT, benchmark-driven evaluation |
| 🎙️ Speech AI | Custom ASR fine-tuning, speaker diarization, streaming pipelines for multilingual and code-switched audio |
| ⚡ Performance | CUDA/Triton kernels, distributed training with FSDP, NCCL over RoCE, inference optimization |
| 🏗️ ML Infra | On-prem GPU clusters, CI/CD for models, privacy-first on-premises deployment |
| 👁️ Computer Vision | YOLO-based detection models for safety and industrial use cases |
ML / Systems
Software Engineering
"Models succeed or fail at the systems layer."
⭐ If something here is useful to you, a star is always appreciated.