A spiking neural network and its ReLU twin, matched in architecture, parameter count, data, seed, and training, compared on DVS128Gesture for accuracy, calibration, robustness under four corruptions, temporal sensitivity, and paper energy. Spiking gives up 3.5 accuracy points and buys a calibration tie, up to +28 points under sensor noise with confidence that tracks its accuracy, and an 18.7x saving on the standard energy accounting. Single seed, n=288, bootstrap CIs on everything, and a live demo.

Natural Language Processing

79.5% win rate

Urdu LLM Fine-Tuning

Qwen 2.5 7B Instruct fine-tuned with QLoRA for Urdu script, Roman Urdu, and Urdu/English code-mixing. The current version wins 79.5% of blind pairwise comparisons against the base model across two independent LLM judges, adds RAG-aware training, and ships as a live Gradio + Modal demo, all for about $60.

Machine Learning & MLOps

73.2% acc · 61% recall

US Visa Approval Prediction

Predicts US PERM labor certification outcomes (73.2% accuracy, 61% denied recall) and explains each prediction with SHAP. A 5-stage MLOps pipeline, GridSearchCV across 5 boosting models, threshold-tuned Gradient Boosting, and a FastAPI backend.

All projects

Writing

Notes from the projects

Write-ups on the projects above, with the numbers and the failures kept in.

Machine Learning · September 3, 2026

Do Spikes Fail Differently? A Matched SNN vs ANN Reliability Study

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 · June 28, 2026

Fine-Tuning an Urdu LLM: Data Curation to Deployment

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 · April 5, 2026

Predicting US Visa Outcomes: EDA to Deployment

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

All posts

Background

Education

Master's in Artificial Intelligence Engineering
COMSATS University Islamabad
2025 - 2027 (expected)
BS Geographic Information Science
University of the Punjab, Lahore
2021 - 2025

More about me

Looking for an ML engineer?

I'm looking for a full-time AI/ML role. If you have one, or a model that needs to reach production, send me a note.

Get in touch

I build ML systems that make it to production

Projects

Do Spikes Fail Differently?