I work across the full machine learning lifecycle: data pipelines, distributed training and finetuning, evaluation, and models running in production. I train and finetune language models, and I build the systems that decide whether one is ready for real users: the infrastructure that serves it, and the evaluation harness that measures whether it behaves as intended. What I find most interesting is the practical side of model safety, getting a model to stay reliably inside the behaviour you specified, and being able to prove it before a failure reaches anyone.
Competitive programmer and side-project maniac by night. I built a dual RTX 5090 workstation so the experiments can run at home.
Toolkit. PyTorch · Hugging Face · vLLM · FSDP-v2 / DeepSpeed ZeRO · LoRA / QLoRA · DPO / GRPO · Axolotl · MLflow / Weights & Biases · AWS (SageMaker, Lambda, Bedrock) · Terraform · SLURM · Docker · Python · C++ · familiar with JAX
A language model for Sudanese Arabic, trained from scratch on my own two-GPU workstation. Trains its own 32k BPE tokenizer, pretrains a 110M-parameter decoder-only transformer (12 layers, 768-dim, RoPE, SwiGLU) on 5B Arabic tokens at 60% MFU, then continues pretraining on Sudanese text and finetunes on dialect chat, with each data decision settled by ablation. Configs and results for every stage are in the README.
Research on where model capacity should go when a model reads a long speech transcript and writes a short structured answer citing the lines that support it. Encoder-decoder and decoder-only models trained under one recipe, up to 3.8B parameters with FSDP2, with every run written up as hypothesis then verdict. At matched size the encoder-decoders lead, and decoder-only models need about twice the parameters to catch up (one seed per cell).
A PyTorch implementation of BERT pretraining and finetuning, written to understand the architecture and the masked-language-modelling objective from the ground up rather than calling into a library.
A multi-task ResNet-34 that wakes on the trigger word "on", then classifies the speaker's gender and the digit (1–9) they spoke. Trigger-word detection trained on SpeechCommands, the classification heads on Speech-MNIST.
A PyTorch implementation of the SRCNN super-resolution model from Dong et al. (2015), alongside a critical analysis of how later state-of-the-art approaches attack the same problem.
A hybrid agent for Pommerman combining rule-based play with Monte Carlo tree search. Against the framework's strongest built-in agent it wins 48.2% and draws 25% — it plays for power-ups and survival, so it reaches the endgame as the strongest player.
A deep Q-network trained with Q-learning to play Atari Breakout from pixels, built to work through value-based reinforcement learning hands-on.
Pong written entirely in NASM assembly and booted in a virtual machine with no operating system, so the code handles keyboard interrupts and draws to the screen itself. Two-player and one-player modes.
A small Unity game built as an excuse to compare search algorithms (BFS, DFS and A*) on a live grid. The shipped version routes the waiter with BFS.
Happy to talk about LLM training, distillation, evaluation, or how any of the projects above were built.