Work
A working portfolio, not a résumé.
Twelve AI systems in production — built, deployed and operated by the AI/ML Unit at Askari Bank. Code we own. Models we can inspect. Infrastructure we run.
Sep 2024 – Present
Askari Bank.
More detailShow less
I built the bank's centralised AI function from a standing start — hiring the team, setting the operating model, and taking the bank's first AI systems into production in-house. The unit is small and in-house by design: we build the models the bank runs, keep the code and weights where we can inspect them, and operate the infrastructure ourselves rather than renting a black box.
That shapes what we ship: production GenAI inside a regulated environment — open-source-first, self-hosted, and instrumented before it scales. Voice banking, cash optimisation, compliance screening, and document work are live; each one earns its place by replacing a capability the bank used to buy.
I also serve as Secretary to the President & CEO's "Bank of the Future" forum — setting the agenda and benchmarking the bank's AI transformation against the digital-first playbooks that reset the standard for the industry.
The discipline is the same one I brought from the lab: measure before you procure, keep data inside the bank's network, and let the numbers — not the vendor — decide.
VP & Unit Head — AI & Machine Learning · Oct 2025 – Present
Senior Data Scientist · Sep 2024 – Oct 2025
A team I hired and lead — twelve AI systems into production, every one built and run in-house. Code we own, models we can inspect, infrastructure we run ourselves.
The programme
VoicePay
Banking for customers who cannot type — Urdu, English and Roman Urdu, 4,200 users at 90% task success
Cash Optimization
PKR 3 billion in surplus branch cash released in a single quarter — idle money put back to work
AI Email Gateway
Consumer Banking correspondence answered within SLA — over 500 emails a day, classified, routed and owned end to end
Conversational assistant
Roughly 1,500 calls a month moved off the phones — 20,000 messages across 8,000 conversations
Compliance suite
168,000 names screened daily at full coverage — trade-based money laundering reporting cut from twenty minutes to under one
Built in-house
Every model the bank's own IP — roughly USD 450K of vendor cost avoided, and no licence fees in perpetuity
Infrastructure & operations
The systems above run on a production AI platform the unit built and operates itself: a self-hosted, containerised setup on Docker infrastructure we manage, with GPU inference on dedicated VRAM held with a service provider. Nothing leaves the bank's network for a model call, and the models stay ours to inspect and retrain.
Capacity was engineered empirically rather than bought to spec. We proved that production load was constrained by system memory, not VRAM, and re-architected accordingly — a decision worth more than PKR 100 million against the alternative.