VP & Unit Head — AI & Machine Learning · Askari Bank
I built Askari Bank's centralised AI function from a standing start — the team, the operating model, and twelve systems in production that the bank owns outright. No vendor dependency, no licence fees, and nothing we cannot inspect.
The portfolio, in production
12
AI systems in production
9
Departments served
1.2M+
Digital customers reached
$450K
Vendor cost avoided
0
Model licence fees
The case
Most banks in emerging markets are buying AI the way they bought core banking in the 2000s — a multi-year contract, a seven-figure license, and a delivery date that slips for two years. It does not work. Traditional vendor deployment cycles in emerging market banking often stall at implementation.
I run the alternative. A small team, hired locally and trained on the actual stack. Open-source models on infrastructure we control. Code that lives in our own repositories, behind our own firewall, instrumented with our own observability. Systems built and shipped for less than the markup on a single vendor PoC — and operated by the same engineers who built them.
The unfashionable claim: in regulated banking, in emerging markets, in-house is not a compromise. It is the only model that ships.
How I run it
Twelve systems shipped in twelve months because the decisions around them were standardised. This is the part that travels.
01
Quantitative use-case scoring replaced subjective High/Medium/Low prioritisation. Executive leadership now uses it for portfolio-level investment decisions.
02
Shared inference, retrieval, and observability rather than a bespoke stack per project. New systems inherit the platform instead of rebuilding it.
03
A standard security and governance review runs before anything ships. No model reaches production without it.
04
Benchmark the real bottleneck before signing anything. That discipline is why the portfolio runs on models we own rather than licences we rent.
Mandate
Governance
Selected work
Systems running in production inside a regulated bank, and the enterprise architecture behind them.
voice banking for the mobile app
Voice-driven fund transfers, balance inquiries, bill payments, and statements in Urdu, English, and Roman Urdu — roughly a 90% task-success rate in production. Built in-house on FastAPI, Whisper, FunctionGemma, and Docker. Proprietary agentic solutions that offset typical six-figure enterprise SaaS licensing fees.
screening, surveillance and variance reporting
Three systems the Compliance function now runs on: trade-based money-laundering variance reporting, cut from twenty minutes to under one; sanctions and watchlist screening at 168,000 names a day with full coverage; and continuous adverse-media surveillance feeding KYC and AML triage. Built on one shared platform rather than three, which is why the third cost a fraction of the first — and why three planned hires were never needed.
Designed and built for a US consultancy
End-to-end design and build of an AWS-native data lakehouse on S3, Glue, Redshift, and Lambda. ISO 27001-aligned. Delivered USD 250,000 in annual savings and a 30% infrastructure cost reduction. The same architecture pattern was reused across several client engagements.
US consultancy · 2023–2024
Deployed for an HR-tech startup
A churn-prediction model for an HR-tech startup, deployed in production. AUC 0.91 on held-out data; downstream the company reported a 35% lift in engagement on targeted retention workflows. Feature engineering and threshold calibration were the load-bearing decisions; the model itself was deliberately simple.
HR-tech startup · 2023–2024
Conversations
Emerging-market banking, in-house AI capability, and the operating model that makes it hold. The fastest path is email.