Developed by Himanshu Yadav | GDG Lucknow APL Hackathon

सुनवाई SunwAI is a premium, resilient, autonomous multi-agent grievance orchestration and resolution ecosystem specifically optimized for Uttar Pradesh's metropolitan municipalities. Built to solve delayed routing, officer fatigue, and language accessibility, SunwAI auto-classifies vernacular complaints (English, Hindi, Hinglish), routes them to the least-loaded officer with predictive SLAs, and autonomously follows up while keeping citizens updated through real-time native SMS messages.

For highly detailed documentation covering the full Problem Statement, Approach, and Cybersecurity methods, please see the docs/ folder in this repository.

This project was built iteratively to tackle the massive bureaucratic bottlenecks present in India's public complaint portals. We successfully evolved the system from a fast MVP (Version 1) to a highly secure, scalable, and intelligent ecosystem (Version 2).

The original version of SunwAI was focused entirely on proving the Multi-Agent Concept.

• Goal: Demonstrate that AI agents can talk to each other and handle basic public routing.
• Tech Stack: FastAPI + basic HTML/JS + Gemini 1.5 Flash.
• Workflow: A user could type a complaint in Hinglish, and a simple linear chain of 4 AI Agents (Intake -> Classify -> Route -> Comm) would process it and output a mocked ticket.
• Limitations: No security. No database. The APIs were completely open, vulnerable to LLM injection, spammers, and malicious DDOS actors. It was highly susceptible to duplicate ticketing (e.g. 5 neighbors reporting the same pothole).

Today, the entire system architecture was rewritten and heavily upgraded for production-grade cybersecurity, web stability, and advanced agentic analysis.

1. Gemini 2.5 Flash Upgrade: Entire backend pipeline migrated from the deprecated gemini-1.5-flash model directly to gemini-2.5-flash, significantly boosting reasoning speed, Hinglish parsing accuracy, and eliminating silent API 404 drops.
2. Citizen Credibility Index (Section 8): Implemented a robust 0-100 credibility scoring algorithm. The AI analyzes if OTP is verified, word count density, geospatial tagging, distress scoring, and duplication flags to assign a priority status (Low, Medium, High, Verified) to prevent malicious or manufactured urgency attacks.
3. OTP & JWT Secure Tunneling (Section 4 & 5):
   • Introduced simulated mobile OTP generation and verification barriers on the frontend before allowing grievance submissions.
   • Replaced flat JSON authentication with JWT tokenized sessions.
   • All API keys safely stripped from the frontend and secured purely in main_secure.py.
4. Prompt Injection & Spam Defense: Introduced PromptSecurityAgent as "Agent 0" to block adversarial LLM injections, and upgraded IntakeAgentV2 with aggressive duplication detection (MD5 hashing) and heuristic spam blockers.
5. Dynamic Incident Clustering: The new IncidentClusterAgent identifies localized micro-outbreaks (e.g., three people reporting a water leak on the same street) and links them to a single parent ticket.
6. Frontend Address Geocoding: Enhanced the public form to capture and inject exact string addresses invisibly into the AI's contextual context layer.

      Citizen Grievance Submission (English / Hindi / Hinglish) + OTP Verification
                                   │
                                   ▼
                      ┌───────────────────────────┐
                      │ 0. Prompt Security Agent  │ Blocks malicious injections & threats
                      └────────────┬──────────────┘
                                   │
                                   ▼
                      ┌───────────────────────────┐
                      │ 1. Intake Agent V2        │ Cleans Hindi dialects, strips spam & dupes
                      └────────────┬──────────────┘
                                   │
                                   ▼
                      ┌───────────────────────────┐
                      │ 2. Classification Agent V2│ LLM extracts categories, tags geo-loc
                      └────────────┬──────────────┘
                                   │
                                   ▼
                      ┌───────────────────────────┐
                      │ 3. Sentiment Agent        │ Evaluates distress levels & tone
                      └────────────┬──────────────┘
                                   │
                                   ▼
                      ┌───────────────────────────┐
                      │ 4. Routing Agent          │ Assigns least-loaded in-charge officer
                      └────────────┬──────────────┘
                                   │
                                   ▼
                      ┌───────────────────────────┐
                      │ 5. Incident Cluster Agent │ Groups similar nearby complaints together
                      └────────────┬──────────────┘
                                   │
                                   ▼
                      ┌───────────────────────────┐
                      │ 6. Communication Agent    │ Compiles warm native Hindi SMS alerts
                      └───────────────────────────┘

Ensure your current directory is the root directory of the project: E:\GDG LUCKNOW\Jansathi-AI.

Open a terminal and execute:

cd backend
# Make sure to activate your virtual environment if not active:
# ..\venv\Scripts\Activate.ps1

# Install requirements (including newly added PyJWT)
pip install -r requirements_v2.txt

# Boot the new secure server
python -m uvicorn main_secure:app --reload --port 8000

The backend server will successfully start listening on http://localhost:8000.

Open a second terminal tab and execute:

cd frontend
python -m http.server 5500

The Web App will run instantly on http://localhost:5500/.

To access the administrative Officer Control Dashboard, a secure glassmorphic gate has been implemented:

• Government Portal Page: Click Officer Panel in the navigation bar.
• Credentials Check:
  • Test Login: Under the new JWT system, the hardcoded admin is admin / admin123.

• Team Name: ATHU
• Developer: Solo; Himanshu Yadav
• Hackathon: GDG Lucknow APL Hackathon 🚀
• Corpus Repository: JanSaathi-AI