Ali Tariq

Designed, trained, and deployed production deep learning models (PPO/APPO multi-agent reinforcement learning) in PyTorch on a Ray/RLlib distributed stack, integrated into AWS cloud services. Outperformed the S&P 500 buy-and-hold benchmark by 5%. Monitored and evaluated model performance in production, diagnosing and eliminating three distinct reward-hacking failure modes through systematic behaviour evaluation, improving reliability and speed. Built and maintained scalable, tested Python data pipelines and APIs (FastAPI), implementing MLOps best practices including CI/CD via GitHub Actions and Docker containerisation. Debugged distributed-training clusters and enforced data-leakage prevention and statistical validation using advanced SQL on noisy financial time-series data.

Built and deployed end-to-end AI-powered systems by integrating ML models and AI agents (PPO/APPO) into production on a Ray/RLlib framework, improving performance over the S&P 500 buy-and-hold benchmark by 5%. Designed AI agent solutions using orchestration (LangGraph, LangChain, MCP), tool use, and RAG, and served them via Python web APIs (FastAPI/Ray Serve) with systematic evaluation gates. Deployed on AWS/GCP using Docker and CI/CD, debugging distributed clusters including shared-state and episode-boundary issues, and enforcing data-leakage prevention. Refactored research code into modular, tested Python packages and translated loosely-specified requirements into working systems, using Claude Code for agentic pair-programming.

Used Python-based EDA to uncover an admissions trend, contributing to a 90% rise in home-student applications. Built Tableau/Power BI dashboards and SQL ETL workflows, translating requirements for non-technical stakeholders. Provided academic support across Mathematics for Computer Science, Information Retrieval, and Computer Security modules.