AI/ML Engineering, Built in Public

Technical deep-dives on building AI/ML systems: LLM fine-tuning, computer vision, NLP, and production ML, with the metrics, trade-offs, and failures behind each project.

Building an Urdu LLM: From Data Curation to Deployment

Machine Learning

June 28, 2026

Building an Urdu LLM: From Data Curation to Deployment

A technical deep-dive into fine-tuning Qwen 2.5 7B for Urdu across three versions: curating a bilingual Urdu/Roman-Urdu corpus, QLoRA training on a rented H100, diagnosing and fixing catastrophic forgetting, building a blinded multi-judge evaluation harness (including a free Claude-Code-CLI judge), fixing the regressions a data change caused, RAG-aware retraining that fixed retrieval's structural failure without turning it into a blanket win, and shipping a live demo on Gradio + Modal.

Building a US Visa Prediction System: From EDA to Deployment

A deep technical dive into building an ML system that predicts PERM labor certification outcomes, covering EDA on 25K records, model selection across 5 boosting algorithms, threshold tuning for class imbalance, SHAP explainability, and deployment on Hugging Face Spaces.

Face Expression Detection: Tackling Class Imbalance with Ensemble Learning and Cloud GPUs

Technical deep-dive into building a facial expression recognition system for group photos. From handling severe class imbalance to cloud GPU training on Modal.com, deploying on Hugging Face Spaces.

Building TeacherRank: A Student-Driven Teacher Review Platform

Technical deep-dive into building a comprehensive teacher rating platform using React, TypeScript, and Supabase. Implementation details for full-stack development, real-time databases, and performance optimization.

Building WaterTrace: ML-Powered Groundwater Prediction Using 22 Years of Satellite Data

Technical deep-dive into building a geospatial AI system for groundwater monitoring in Pakistan. From processing NASA satellite data to deploying time-series ML models achieving R²=0.89 accuracy.

12 min read

Machine Learning+7