I build minds for machines.

Problem: Pakistani clothing brands lose sales to unanswered Instagram and WhatsApp DMs, and to a 30–35% Cash-on-Delivery return rate driven by unconfirmed orders. Before building, our squad validated this on the ground — messaging 16 Pakistani clothing brands as real customers within one hour to quantify the unanswered-DM problem, then interviewing brand owners to shape the product.

Solution: An agentic AI copilot that reads customer questions in Urdu and Roman Urdu across WhatsApp and Instagram DMs and replies instantly with catalog-grounded answers on sizing, delivery, and COD confirmation — with a one-tap handoff to a human and live stock management from a React dashboard.

Context: Built in a 4-person squad during Comebck Pakistan Cohort 1 — an 8-week, YC-style accelerator and the first Pakistani product accelerator. Selected as a Top 5 project of the cohort.

Technology: Agentic AI and LLM integration, React dashboard, Python/FastAPI backend, LangChain, Redis, and Docker for a fully containerised stack.

Problem: Long-lived systems accumulate state snapshots faster than they can be stored, but discarding data uniformly loses the records that actually carry information.

Solution: An ETL pipeline that computes Shannon entropy for each record via automated database triggers, then routes it through rule-based preservation logic — discard, compress, preserve, or archive — across a normalized (3NF) 10+ table schema. A real-time PHP dashboard surfaces entropy trends and audit logs.

Technology: PostgreSQL with PL/pgSQL functions, triggers and views; Python; PHP dashboard with Bootstrap; Shell tooling; Docker Compose for a one-command environment.

What it covered: A two-month intensive spanning classical machine learning, deep learning and NLP, agentic AI with LangChain and RAG, and MLOps tooling (Git, DVC, MLflow) end to end.

Curriculum arc: mathematical foundations for modern AI (linear algebra, calculus, probability, information theory) → NLP and transformer architectures including self-attention, BERT and GPT with hands-on fine-tuning → post-training and alignment (instruction tuning, RLHF/RLAIF, PEFT, and Chain/Tree/Graph-of-Thought reasoning) → agentic AI with ReAct, tool-using agents, memory and multi-agent orchestration, plus advanced RAG → inference and deployment (ML/LLMOps: versioning, CI/CD, serving, observability, cost and security) → a capstone building and deploying an autonomous agent system.