Data & AI engineer. I work on model efficiency — pruning, distillation, low-rank adaptation — and on the enterprise data platforms that put models into production.

Two tracks, deliberately: the research is where I go deep, the platform work is what I ship day to day.

Making large models smaller, cheaper to adapt, and still worth running.

BaCP — Backbone Contrastive Pruning

Prunes a pretrained network to 99% sparsity while preserving its representations, using a four-term contrastive objective (PrC / SnC / FiC + CE). Benchmarked against magnitude pruning, SNIP-it and WANDA at 0.95 / 0.97 / 0.99 sparsity across ResNet, VGG, ViT, DistilBERT and RoBERTa.

Every equation the code computes is derived to its source in docs/foundation.md, with a full citation audit beside it, and the invariants are covered by a pytest suite.

LoRA — Low-rank adaptation, from scratch

Why fine-tuning updates have low intrinsic rank, and what falls out of that — rank decomposition injected into frozen weights, derived and implemented end to end.

DPO — Direct Preference Optimization

Preference alignment without reinforcement learning. Built on SmolLM from first principles: no reward model, no PPO, just log-probability ratios between chosen and rejected responses.

Warehouses, pipelines and the movement of data at enterprise scale. This is the day job.

azure-lakehouse-patterns — change data capture, Delta Live Tables pipelines and ETL patterns on Azure Databricks.

sql-warehouse-medallion — a dimensional warehouse built from scratch in T-SQL, bronze through gold.

LUMSegmentation — semantic and instance segmentation of field-collected plant imagery. A U-Net with LSTM units and an optimised Mask R-CNN, trained on data I cleaned and annotated by hand. LUMS Directed Research Project.

StructureFromMotion — 2D-to-3D reconstruction on the heritage-recon dataset, in OpenCV.

ml-from-scratch — backpropagation, knowledge distillation and regression, implemented without frameworks.

Python · PyTorch · T-SQL · TypeScript

Azure Databricks · Delta Live Tables · Power BI · Next.js · Prisma

OpenCV · scikit-learn · pandas