M.Sc. student in Electrical Engineering and Computer Science, and previously an Applied Science Intern at Microsoft. I work on machine learning systems and the evaluation infrastructure around them — the part that tells you whether a model actually does what you claim it does.
Most of what I build falls into three groups: interpretability and safety research on language models, retrieval and representation learning, and performance work close to the hardware.
| Project | What it is |
|---|---|
| DPO safety representations | Does DPO build a new safety representation, or amplify one the base model already had? Probing and steering a refusal direction across a Base → SFT → DPO chain. |
| Attention mechanism paradigms | Attention on 12 execution substrates and 17 implementations, measured and modelled stage by stage. Softmax is 0.4% of the FLOPs and up to 35% of the measured runtime; a silent float64 promotion under NumPy 2 made FP32 look 20× slower. |
| Optimization methods | Six optimisation problems from scratch in NumPy — metaheuristics, derivative-free fitting, and backpropagation raced against a genetic algorithm. |
| Code search engine | Semantic code search over CoSQA. A bi-encoder fine-tuned with InfoNCE lifts MRR@10 from 0.79 to 0.91. |
| Written text recognizer | Handwritten paragraph recognition, ResNet encoder into a Transformer decoder, served over FastAPI. |
Python · PyTorch · PyTorch Lightning · NumPy · JAX · scikit-learn · FAISS · FastAPI · Docker · CUDA · MATLAB