About this project
An agent system built for MTM Hackathon '26 (Mind-to-Machine), GDG On Campus CUI Wah, by Team AXIONS of GDG On Campus CUI Attock.
The task
Read a pack of about ten messages - emails, WhatsApp texts and notices, in English and Roman Urdu, some irrelevant and some out of date - and plan a four-hour physical hackathon: venue, schedule, food, judging and budget, respecting every stated constraint. At judging, the pack is unseen: different numbers, different wording, different senders.
The principle
The model reads; the code decides. Language models are good at reading messy human messages and bad at arithmetic you must trust. So the model does exactly one job - turning message text into structured, dated facts - and plain code does everything else: merging, planning, the six validity checks, rejecting harmful updates, and computing every risk and recommendation. Every figure in a plan can be traced to the message it came from.
How it is built
- One small model - openai/gpt-oss-20b via Groq, for every call. A full run takes about five seconds and one reading call; an update takes at most two calls.
- A Python core (standard library only) that runs the whole system from one command - and this website, whose planner is a line-by-line port of the same engine.
- A guarded fact base - facts carry their source message and date; later quotes supersede earlier ones; quantity conflicts are caught by a deterministic scan in digits, spelled-out words and Urdu digits.
- Proof, not promises - an offline test suite (23 checks) covers the defences and the end-to-end flows, and every plan ships with its full trace.
The team
Team AXIONS - GDG On Campus CUI Attock. Lead: Abdul Wahid. The system was designed so that any member can explain any part of it: five parts, each with one job, each missed if removed.