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 four groups: LLM evaluation and the data behind it, post-training and model internals, retrieval and representation learning, and performance work across hardware and computing paradigms.
| Project | What it is |
|---|---|
| DPO safety representations | How safety SFT and DPO change a small model's internals: SFT preserves the harmful–benign contrast, DPO redirects it. Matched starts, three runs per arm, steering and ablation tests. Paper accepted at TELFOR 2026. |
| 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