Write-ups on LLM fine-tuning, computer vision, NLP, and production ML, with the numbers and the failures kept in.
Machine Learning ·
A spiking neural network and its ReLU twin, identical in everything but the neuron, compared on an event-camera gesture dataset: accuracy, calibration, four corruptions, a timing control, and paper energy. Spiking costs 3.5 points and buys honest confidence under noise, a calibration tie, and 19x on paper. With a live demo you can break yourself.
14 min read
Machine Learning ·
Fine-tuning Qwen 2.5 7B for Urdu across three versions: a bilingual Urdu/Roman-Urdu corpus, QLoRA on a rented H100, catastrophic forgetting diagnosed and fixed, a blinded multi-judge evaluation harness, RAG-aware retraining, and a live Gradio + Modal demo, all for about $60.
20 min read
Machine Learning ·
An ML system that predicts PERM labor-certification outcomes: EDA on 25K records, five boosting algorithms compared, threshold tuning for class imbalance, SHAP explainability, and deployment on Hugging Face Spaces.
15 min read
Computer Vision ·
A facial expression recognition system for group photos: severe class imbalance handled with ensembles, cloud GPU training on Modal.com, and deployment on Hugging Face Spaces.
15 min read
Web Development ·
React, TypeScript, and Supabase in production: multi-dimensional teacher ratings, real-time sync, and the route-splitting pass that cut the initial bundle 60%.
10 min read
Machine Learning ·
A geospatial AI system for groundwater monitoring in Pakistan: NASA satellite data processed into time-series ML models that reach R²=0.89, served through district-level maps.
12 min read