Civic AI
AI-powered civic issue reporting and government management platform that won Champion at the AI & API Hackathon 2026.
Full-Stack Architect & AI Lead
Team of 4 (Error 429: Too Many Devs)
Hackathon Champion Project (Team Error 429: Too Many Devs)
AI & Full-Stack Platform
August 2026
1st Place Champion
The Problem & Real-World Friction
Citizens in municipal areas face severe communication friction when reporting civic issues in mixed Banglish/English. City authorities lack automated triage tools, resulting in backlogs, missed emergencies, duplicate complaints, and exposed citizen PII.
Architectural & Technical Approach
Architected an end-to-end civic reporting platform with multimodal LLM classification (Banglish & English), automated severity categorization, resilient Gemini primary / Groq fallback failover, Supabase Row Level Security for strict PII isolation, and multi-signal duplicate detection combining semantic text embeddings, geospatial proximity (PostGIS), and temporal windows.
Key Technical Highlights
Multimodal AI Municipal Triage System
Banglish & English AI Triage
Multimodal LLM pipeline classifying municipal citizen reports in conversational Banglish and formal English with automatic priority severity scoring.
Zero PII Citizen Privacy
Strict privacy safeguards with Supabase Row Level Security (RLS) guaranteeing that personal identifying data is never exposed in public triage feeds.
Multi-Signal Duplicate Detection
Combines semantic vector embeddings (pgvector) with geospatial proximity and temporal clustering to eliminate duplicate municipal work tickets.
Resilient Dual-AI Failover
Dual LLM provider pipeline with Google Gemini as primary inference engine and Groq / Llama as automatic timeout-triggered fallback.
Project Results & Impact
Won 1st Place Champion at AI & API Hackathon 2026 – First Edition (IEEE CS SEU). Reduced complaint triage time from hours to seconds.
Reproducible Benchmarks & Test Setup
Test setup, evaluation procedure, and validated outcomes reported for this architecture:
April 2026 (AI & API Hackathon 2026)
Automated batch harness testing 120 synthetic municipal reports (60 Banglish, 40 English, 20 corrupted/vague).
Baseline: Manual civic complaint triage avg 2–4 hours response time.
Outcome: Automated AI triage completed in 1.4s avg (p95: 2.1s). 94.2% classification accuracy against curated ground truth.
Separation of Implemented, Prototyped & Planned Features
- •Multimodal Banglish and English issue classification
- •Automated priority scoring and severity categorization
- •Supabase RLS citizen privacy data isolation
- •PostGIS geospatial radius clustering (50m proximity window)
- •Dual-provider automatic failover with timeout-triggered fallback
- •Citizen automated WhatsApp notification webhook
- •Municipal contractor field technician work-order dispatch SMS bridge
Role & Contributions
Full-Stack Architect & AI Lead
Team of 4 (Error 429: Too Many Devs)
- Architected dual-provider LLM failover pipeline (Gemini 1.5 Flash primary with Groq Llama 3 70B fallback).
- Engineered multi-signal duplicate triage engine combining pgvector semantic search (1536-dim) with PostGIS radius clustering.
- Configured Supabase Row Level Security (RLS) policies to cryptographically isolate citizen PII from public feeds.
- Designed responsive Next.js 15 interface and incident dispatch dashboard.
Technologies Used
10 toolsGreen Life Society
Production full-stack web platform for an environmental nonprofit organization.
View next case studyInterested in this architecture?
Have a question about the system design, tech stack, or potential collaboration?
Let's discuss