I build ML systems end-to-end: feature engineering, model selection, explainability, and the cloud infrastructure that keeps them running in production. Based in Pakistan, focused on problems at the boundary of statistical modeling and real software engineering. Current direction: deep learning architectures, production MLOps and LLM-based reasoning systems.
▸ Neural architectures from scratch implementing forward pass, manual backprop
and optimization in PyTorch at the layer level before touching nn.Module
▸ Production model deployment containerized ML pipelines with versioning and monitoring baked in from the start; treating a trained model and a deployed model as different engineering problems
▸ LLM-based reasoning systems building applications where language models function as reasoning components inside larger pipelines, including fine-tuning workflows for domain-specific performance
ChurnIQ Predicts customer churn using XGBoost with SMOTE for class imbalance, SHAP for per-customer plain English driver explanations and a lightweight exponential survival model that estimates how long each customer is likely to stay with an 80% confidence interval. Deployed as a full Streamlit app with snapshot diffing, industry benchmarks, and one-click PDF reports.
AI Recruitment Agent Screens PDF resumes against job descriptions via Gemini API analysis, then passes scores through a deterministic decision engine that is fully decoupled from the AI narrative, Score > 85 shortlists, 60-85 escalates to HR review, below 60 rejects. Pydantic validates every structured JSON response before it reaches the UI, preventing silent failures on malformed model output.
LR(0) Parser Full LR(0) construction pipeline from the Dragon Book, closure, GOTO, canonical collection, ACTION/GOTO table, and SLR(1) refinement for reduce actions. Ships with a 6 tab tkinter GUI including step through parse animation and DFA export, conflict detection highlighted in the table and 13 unit tests covering grammar parsing and conflict cases.
Kubernetes ECR Demo Complete Docker → ECR → Kubernetes pipeline on a single EC2 t3.micro. Uses an IAM role for ECR auth instead of long lived keys on the instance, socat to bridge the Docker driver's NodePort to the host interface and a 2 GB swap file to keep Minikube stable on 1 GB RAM, problems you only encounter by actually building it.
Tier 1 Used in real projects:
Tier 2 Actively building with: