EXPERIMENT LOG
07 experiments · one of them running a real business.
Each run carries its own heat signature — hotter means fresher on the ramp.
A headless ERPNext v16 platform running a real pharma distribution business — order to cash, role-shaped workspaces, and an LLM analyst that reviews the day's events and reports what looks wrong.
Automated land-price heatmap of Islamabad — spatial interpolation, choropleth layers, and a data-storytelling pass that makes the market legible.
End-to-end MLOps platform predicting molecular toxicity — experiment tracking, model registry, and containerized deployment.
Airflow-orchestrated training pipeline with MLflow experiment tracking — repeatable, scheduled, observable ML delivery.
Comparative fine-tuning study — full training under DeepSpeed against LoRA adapters, with cost, speed, and quality trade-offs measured.
Productivity browser extension with AI-assisted right-click workflows — quicker actions, cleaner UX.
Production-style authentication system — secure user flows, session handling, and backend integration.
CAPABILITY MATRIX
RAG, classification, recommendation systems, vision experiments, and LLM workflows.
Trackers, heatmaps, explorers, scraping systems, and analytics-first interfaces.
Airflow, MLflow, Docker, CI/CD, and reproducible workflows for real deployment.
Browser extensions, auth flows, and frontend layouts that make work feel finished.
TRAINING NOTES
Most models die in notebooks. I build the pipeline that gets them out — tracked in MLflow, shipped in Docker, orchestrated in Airflow, and wrapped in interfaces people actually want to use.
FINAL EPOCH