I build multi-agent and RAG systems: citation-grounded retrieval, agent pipelines, and the deterministic checks that keep an LLM honest.
Before that, 8 years in real-time systems and ML, including predictive models used on Star Sports’ live IPL coverage.
Hi! I’m Ali, an AI engineer. I build multi-agent and RAG systems: retrieval-grounded chatbots that cite real sources instead of making things up, and agent pipelines that turn dense financial documents into structured, source-linked analysis for a private equity firm. The hard part is never getting a model to talk. It’s getting it to be right, and proving it.
Before AI, 8 years in real-time systems and ML. I built predictive models on bowler and team data every IPL season and designed the ball-by-ball pipeline behind them, with the analysis going out live on Star Sports to over 200 million viewers. After that I led the ML pipeline for real-time body tracking that runs entirely on low-end Android phones. Same discipline either way: hard latency budgets, and a number that has to be right.

An MCP server for ATS resume scoring with zero human-facing UI. One tool, analyze_resume, meant to be called by an AI agent, not a person.

Four historical figures, one citation standard. Gandhi, Jinnah, Einstein, and Darwin, each grounded in their real writings, every claim cited.

An AI that speaks as Muhammad Ali Jinnah, grounded in 10,000+ chunks of his actual speeches, letters, and correspondence.

An app that makes up facts and citations, deadpan, on purpose. A joke about how confident LLMs sound when they’re wrong.

A production pipeline that reads investment documents for hedge funds and PE firms and produces structured analysis.

Predictive models on bowler and team data, built each IPL season to decode game plans, with the output going out live on Star Sports.

A neon endless runner where you flip gravity instead of jumping. Ten zones, 30+ skins, and no patience for you dying slowly.

A fruit-slashing game built on the pose-tracking toolkit. Your hands are the swords, your body is the controller.

A mobile cricket game that tracks your bat and body in real time using just the front camera. No depth sensor, no extra hardware.
Took the cricket tracking tech and generalized it into a toolkit for building any pose-tracking game. Used it to make baseball and tennis titles so far.

An open-source Unity package that drives any Humanoid-rigged avatar from a plain webcam, no depth camera, no motion capture suit.

A VR cricket segment for live IPL broadcast on Star Sports: real match data piped through Kafka, so anyone in studio could play a real delivery 30 seconds after it was bowled.

A public VR cricket activation. 80,000 fans played real deliveries on HotstarVR across India, and 5,000 more at the Caribbean Premier League.

A standalone VR cricket game for the Meta Quest 2. No studio, no broadcast truck, just the headset.
Built with Synapse: a real cricket bat with submillimeter SteamVR tracking and force feedback on every hit, for a fully immersive batting experience.

Multiplayer engineering work on Go, Operative!

Built the first MVP of Tower’s Gate for Hayabusa Studios. Inspired by Asheron’s Call, the cult classic nobody but us seems to remember.

A city-builder in the Homescapes mold: minigames, a currency economy, property flipping. Led a team of four to build it in six months.

One of my first projects. A wrestling federation management game for Checkmate Creative, built for mobile.

A mixed reality documentary with interactive holograms of real Pakistani women sharing their stories of gender disparity.
Multi-agent and RAG systems for deal analysis at a private equity firm: source-linked memos, a citation-grounded analyst chatbot, and agents that pull historical insight out of company financial reports. Runs on the client’s own servers with guardrails and prompt-injection defense.
Onboarding and campaign creation built as LangGraph agents, every site action reachable through chat via tool calls, and tiered content moderation across text, documents, and video.
Cited RAG across Jinnah, Gandhi, Einstein, and Darwin, each on an isolated corpus with its own citation trail.
Historically-grounded AI persona of Jinnah, built on a hybrid RAG engine over the Jinnah Papers corpus. Runs on Cloudflare Workers.
Led ML and Unity engineers building real-time 6DoF bat and body tracking on low-end Android phones, no depth sensor. Owned the full ML pipeline: synthetic dataset creation, annotation, training, and on-device inference.
Predictive ML models on bowler and team analytics each IPL season, live broadcast analytics for 200M+ viewers, and the ball-by-ball data pipeline behind them. Also led the real-time cricket simulation for Star Sports India.
Led a team of four to deliver in six months, cutting timelines by about 20%, and built real-time multiplayer.
Multiplayer VR with LEAP Motion hand tracking, held to a stable 90 FPS.
Private equity has spent decades scrutinizing the model and waving the argument underneath it through. Large language models are the first thing capable of testing that reasoning with the same rigor, and here's what that looks like in practice.
Claude-e-Azam is a chatbot that answers as Muhammad Ali Jinnah, grounded entirely in his real historical record, with a citation on every real claim. Here's how it actually works.
How to take MediaPipe's 3D body landmarks from a plain webcam and use them to animate a full Humanoid-rigged avatar in Unity, running entirely on-device with no server and no motion capture suit.