Work
FrontendJan 2026

Leafline

Civic Issue Reporter

ReactNode.jsGroq APIPostgreSQLTailwind CSS
Read Post-MortemSource Code
01 / 05
Leafline screenshot 1
// overview

// Overview

Leafline is the original proof-of-concept that inspired JanSamadhan. Built in 24 hours for the InnovateNSUT Hackathon, Leafline parses photos of municipal issues (e.g. potholes, broken street lights), resolves GPS coordinates, and runs descriptions and image labels through high-speed LLM JSON interfaces to extract damage severity and route details directly to repairing departments. It won 1st Place out of dozens of competing engineering projects.

// The Hackathon Challenge

During the hackathon, the core challenge was to build a system that could intelligently classify and route arbitrary civic reports without needing months of training data for a custom ML classifier. We wanted to build something that was immediately functional for a city manager, with zero warm-up time.

To make the MVP viable under extreme time constraints, we decided to leverage generative vision models but faced a problem: typical cloud model API latency (e.g. GPT-4V) took anywhere from 2 to 5 seconds per request. This would make live judging and real-time dashboard interaction extremely sluggish.

// Technical Solutions & Architecture

#

// 1. Ultra-Low Latency Inference with Groq

  • Integrated the Groq API running Llama-3-Vision models to achieve near-instantaneous JSON inference.
  • By configuring strict system instructions and formatting templates, we forced the model to return structured JSON payloads under 200ms.
  • The payload contains:

    • ```json

    {

    "category": "road_damage",

    "severity": "high",

    "required_equipment": ["asphalt", "roller"]

    }

    ```

    #

    // 2. Taxonomy-Guided Dynamic Routing

  • Rather than training a neural net to classify reports, we established a strict department taxonomy in the LLM's prompt context.
  • The prompt matches keywords and spatial contexts to one of 6 default public departments (Sanitation, Roadways, Power, Forestry, Water, and Police).
  • This approach resulted in a 100% routing accuracy during the judging round, handling messy user typos and conversational complaints with ease.

    • #

    // 3. Interactive Web Dashboard

  • Deployed a single-page React client coupled with Tailwind CSS for layout.
  • Integrates a mapping module that places coordinates on a live interface, letting municipal workers see red (critical), yellow (moderate), and green (minor) pins representing public hazards in real-time.

    • // Key Takeaways

    Leafline proved that LLMs could perform high-velocity classification tasking on civic data. It became the technological bedrock for JanSamadhan, leading to a live Delhi ward pilot deployment.

    // metrics
    24-hour hackathon MVP
    1st Place Winner
    100% routing accuracy
    // stack
    ReactNode.jsGroq APIPostgreSQLTailwind CSS
    // challenges & solutions
    // 01

    24-Hour Scoping Constraint

    Building a fully functional reporter, dashboard, and routing engine within a day. Chose the Groq API for near-instantaneous JSON inference to ensure real-time responsiveness during judging.

    // 02

    Smart Routing without ML Models

    Categorizing and routing complaints dynamically without training a custom machine learning classifier. Solved by configuring Groq LLMs with structured prompts and strict department taxonomies.