Machine Learning
WhyRisk
A live explainable healthcare risk platform: a LightGBM diabetes risk model trained on 70,692 health records (BRFSS), with patient‑level SHAP explanations and role‑based access for patients, doctors, and admins.
Case study, Applied AI
An agent system built in one day at MTM Hackathon '26 that reads a pack of scattered messages — emails, WhatsApp texts and notices in English and Roman Urdu — and produces a complete, validated plan for a 4-hour hackathon: venue, schedule, food, judging and budget. The language model only reads; every decision and figure is computed and checked in code.
Context
Event details arrive scattered across messages that contradict and update each other. A useful plan has to reconcile them, respect every hard constraint, and stay valid when a plausible-sounding message is wrong.
My role: Team lead and presenter, Team AXIONS (with Abdul Moiz and Jawad Hassan), GDG On Campus CUI Attock.
Implementation
One extraction call turns raw message text into structured facts, each with its source message and date. A later message from the same sender supersedes the earlier one, and item names are matched flexibly so a correction replaces an entry instead of double-counting it.
Teams, judges needed, people to feed, the cheapest venue that fits, the food tier, a schedule that sums exactly to the event length, and the full budget are computed in plain code, so every figure is exact and auditable.
Six validation checks run after every build. An update that breaks a previously passing check — or imposes a quantity below a hard requirement — is rejected, the plan stays valid, and the reason is printed in the trace.
Evidence
The points below summarize the public repository and owner-provided CV where noted. Read the source for implementation details and the evaluation setup before drawing further conclusions.
Plan produced from the brief's sample pack (177 people, Hall A); all six validation checks pass.
Full run against the brief's 2-minute limit; updates use at most 2 model calls against a limit of 5.
Regression suite runs against a stub model: sample pack, budget change, the same change twice, and the plausible-but-wrong message (rejected, plan unchanged).
One small model for every call, as the brief's small-model bonus required.
Boundaries
Continue exploring