Data scientist and ML engineer. I build models that survive contact with real data, real regulators and real users — then document them so your team can run them without me.

I'm a data scientist and ML engineer with two years building production machine learning inside regulated financial services — an environment where the model is maybe 20% of the job and the rest is data lineage, monitoring, retraining, and being able to explain every decision to a risk committee.

I currently work in-house at a multinational bank, where I own systems like:

• A multi-country early-warning system that flags customers heading into low-balance stress — per-market configuration, champion-aware retraining, capacity-based alert thresholds, rebuilt from scattered notebooks into a config-driven Python library.
• LangGraph multi-agent and hybrid RAG pipelines for internal knowledge and conversational banking, with hard isolation boundaries so one user can never retrieve another's data.
• The MLOps around all of it: Azure ML, MLflow, Airflow, Docker.

Before that I worked at Standard Chartered. I hold a BSCS from IBA Karachi.

Alongside client work I do adversarial ML research — most recently several thousand attack runs studying graph injection attacks against financial fraud detectors, and what realistic domain constraints actually cost an attacker. This keeps me honest about failure modes. I once traced a "95% answer faithfulness" metric back to plain unigram overlap; it wasn't measuring groundedness at all. If a number in your pipeline looks too good, I'll tell you why before you present it.

I also build and ship my own products — a made-to-order poster storefront, an interactive game-theory app, a daily puzzle game — so I'm comfortable owning something end to end: data, model, API, frontend, deployment.

Good fits for me: fraud, risk and credit models; churn and early-warning systems; RAG and agent pipelines that need real evaluation rather than vibes; cleaning up research code into something deployable; analytics and forecasting for finance and fintech.

What you get: honest scoping — I'll say so if you don't actually need ML; written updates you don't have to chase; documentation and code your team can still run six months after I'm gone. I raise problems early rather than at the deadline.

I spent a summer teaching Python and applied AI to complete beginners, which is roughly why I can explain a model to a non-technical stakeholder without either dumbing it down or hiding behind jargon.

Outside work: weightlifting, cooking, and reading far too many papers.

Work Terms

Availability
Based in Karachi, Pakistan (PKT, UTC+5). I take on 15–20 hours per week of freelance work and keep a hard cap so active projects don't slip. Core overlap: 6am–1pm UK, 1pm–8pm Singapore, and early mornings US Eastern by arrangement. Scheduled calls on evenings PKT and weekends.

Communication
Guru WorkRoom is my primary channel — it keeps scope, files and decisions in one auditable place. I reply within 24 hours on weekdays. I work async by default and send a written progress update at least twice a week, plus a short summary at every milestone. Video calls for kickoff, scope changes and final handover; I'll always follow a call with written notes so nothing lives only in someone's memory.

Payment
All work through Guru SafePay. Fixed-price projects are broken into milestones with the first funded before I start; hourly work is tracked and invoiced weekly. Minimum engagement 10 hours. Rate on request and quoted per project — I'll give you a fixed number after a short scoping call rather than a vague range up front.

Scope and revisions
Every project starts with a written scope: deliverables, data, acceptance criteria, timeline. Two rounds of revision are included within that scope. New requirements are welcome but become a new milestone rather than silent extra work.

Data and confidentiality
Happy to sign NDAs. I work with anonymised, synthetic or sample data wherever possible, and I don't take custody of production credentials — I work inside your environment, under your accounts and your controls. Full IP and code ownership transfers to you on final payment.

Not a fit
Unpaid trial tasks, guaranteed model accuracy before I've seen the data, or work requiring me to misrepresent results.