What I Learned Winning InnovateNSUT '26 in 24 Hours

Behind-the-scenes engineering decisions and tactical hackathon strategies that turned a raw idea into a 1st place civic issue reporting platform.

// summary (tldr)

Focus on speed-to-judgement: why API response times and visual indicators are more important than complex ML models during 24-hour sprints.

Sprinting from Zero to One

Hackathons are not about writing perfect, production-grade clean code. They are about rapid validation, high-velocity prototyping, and building a product that can be clearly demonstrated to judges in a three-minute window.

At InnovateNSUT '26, we competed against dozens of engineering teams to build a solution for local city issues. We built Leafline, an automated civic issue reporter that processes pictures of public hazards, categorizes them, and maps them to government response squads.

In this post, I will share the system design shortcuts and presentation strategies that helped us win 1st Place.

---

1. Optimize for Speed-to-Judgement

The biggest mistake hackathon teams make is spending 20 hours building a highly complex, custom ML model that performs slightly better but takes forever to run. During live judging, a judge will test your app once or twice. If the loader spins for more than three seconds, the judge assumes your app is broken.

We optimized for immediate feedback:

Instead of setting up a local PyTorch model on a slow CPU, we integrated the Groq API running Llama-3-Vision.

Because Groq runs inference on ultra-fast hardware, our image-to-JSON classification took under 200ms.

When the judge uploaded a photo of a pothole, the category and severity tags popped up almost instantly, creating a powerful "wow" effect.

---

2. Dynamic Visual Hierarchy

A clean, responsive dashboard that updates in real-time is essential. We built our client using React and styled it with Tailwind CSS:

We used a clean dark theme styled with warm accent colors to keep visual noise low.

We mapped coordinates onto a mock city layout with pulsing map pins. The pins were color-coded by severity: red for emergency hazards, yellow for medium issues, and green for minor reports.

This made the scope and impact of the platform immediately obvious to judges looking at the dashboard screen.

---

3. Strict Prompt Scoping

When relying on LLMs for classification, model hallucinations can derail your live demo. If a user uploads a photo of a broken street light and the model suggests "call the fire department," the demo fails.

We wrote strict, structured prompts using JSON schema definitions:

{
  "department": "Power & Electricity",
  "priority": "medium",
  "short_summary": "Flickering street lamp reported."
}

By constraining the model to output standard enum options (only mapping to our six defined departments), we guaranteed 100% routing consistency during the live judging tests.

---

Conclusion

To win a 24-hour hackathon, focus on high-speed inference, clean visual states, and reliable prompts. Leafline's success laid the foundation for JanSamadhan, which is now a live municipal pilot in Delhi.