I build AI that has to work outside a demo: voice assistants on government kiosks, models small enough to run offline, and fraud detection for banks. I train the models and run the GPUs myself.
Voice agents that act
Speech-to-speech agent, 20+ tools, self-hosted avatar
I write the code, train the models and run the servers myself.
Speech-to-speech assistants that call real tools: sign-ins, queue tokens, record lookups, payments.
LoRA and QLoRA on Apple MLX and NVIDIA CUDA. Big models teach small ones that run offline.
Self-hosted lip-sync streamed over WebRTC at 25 fps. It stops talking when you interrupt it.
NVIDIA T4, L4 and L40S, in the cloud or on a server with no internet access.
Test suites of a thousand cases, judged by a model. A checkpoint that regresses does not ship.
AML and fraud models whose reports a person signs off before anything is filed.
Featured work
Every number below comes from the work itself. Each one links to the full case study.
People ask out loud, in Urdu or English. It answers from the official catalogue, issues queue tokens and looks up records, and an avatar I host myself speaks the reply.
Gemini taught two small models: a LoRA-tuned Qwen3-1.7B that picks verified answer clips, and a fine-tuned Whisper. Nothing touches the internet at run time.
A Gemini agent that answers texts and voice notes in nine languages and scripts, files complaints with a reference number, and never drops a message during a deploy.
Graph models find laundering patterns. A self-hosted LLM drafts each report from cited evidence, and a person signs it before anything is filed.
Roughly the same four steps on every AI project.
I sit with the people who will use it and collect real examples. Those examples become the first test set.
A hosted model to prove the idea, then a smaller one I can run myself if cost, privacy or offline use calls for it.
A working demo in the first days, then a run against the test set at the end of each one.
Every release is scored first. If a new version is worse at something the old one did well, it waits.
About
I lead the AI Wing at the National IT Board in Islamabad. Before that I ran engineering at a FinTech AI startup in Calgary and an EdTech company in Riyadh, and spent six years on healthcare and distributed systems software. Eight years in, I still write most of the code myself.
Tell me what it needs to do. I'll tell you how I would build it.
Book a 30-minute call
