Post-Mortem: Leafline

During the InnovateNSUT hackathon, we wanted to build a zero-friction reporter. Users shouldn't need to fill out a long form to report a civic issue. They should just snap a picture and have the system do the rest. The key problem was extracting structured data (category, coordinates, severity) from arbitrary images and text in under a second to pitch a working demo to the judges.

I architected the Express.js backend and LLM ingestion pipeline. I set up the intake flow to feed incoming description strings and metadata to the Groq API. It parsed user intent into a clean JSON output matching our database taxonomy, allowing immediate database writes.

We had 24 hours and a $0 budget. More importantly, judge demos happen in seconds. Groq’s Llama inference delivered token output in under 200ms, which allowed us to showcase real-time guided chatbot chats without any "loading..." spinners during the judging rounds. It was the differentiator that made the app feel fast and production-ready.

Instead of hardcoding the department categories inside the system prompt, I would implement a simple configuration schema that maps categories dynamically, which would have made scaling the dashboard to new departments trivial.

- Speed is a feature. Near-instant LLM response times completely change the user experience and feel much more premium. - Venue network limits are the ultimate hackathon constraint. Always build offline fallbacks or heavy clientside compression. - Under time pressure, a simple Node/Express stack beats complex frameworks every time due to setup speed.