PRAGATI

Progressive Reasoning Agent for Global Advanced Total Intelligence

India's Self-Assembling Public Health Intelligence System

Built for Google Gen AI Academy APAC Cohort 1 Hackathon | Track 1: AI Agents with ADK

Live Demo: https://pragati-366193575719.us-central1.run.app

---

What Makes PRAGATI Unique

Most AI agents are hand-coded with fixed tools. PRAGATI self-assembles its own MCP tools at runtime.

At boot, PRAGATI:

1. Cartographer introspects the AlloyDB schema -- no hardcoded table names
2. Forge dynamically generates MCP Toolbox tool definitions from the discovered schema
3. Root Orchestrator (ADK agent) routes natural language queries to the right tool
4. Gemini Flash synthesizes a natural language answer from the raw data

Add a new table to AlloyDB, restart PRAGATI, and it automatically has a new tool. Zero code changes.

---

Architecture

User Query (Natural Language)
    |
    v
Root Orchestrator (Google ADK Agent)
    |
    +---> Cartographer ---> AlloyDB introspect_tables()
    |                           |
    |                           v
    +---> Forge ----------> MCP Tool Registry (self-assembled)
                                |
                    +-----------+-----------+
                    v           v           v
              query_health  query_      query_
              _indicators   facilities  disease_reports
                    |
                    v
                AlloyDB (PostgreSQL + pgvector)
                    |
                    v
              Gemini synthesis ---> Answer + Data Table

---

Stack

Component	Technology
LLM	Gemini Flash (Vertex AI)
Agent Framework	Google ADK
Tool Protocol	MCP Toolbox for Databases
Database	AlloyDB (PostgreSQL + pgvector)
Embeddings	Vertex AI text-embedding-004
API	FastAPI + asyncpg
Deployment	Cloud Run (us-central1)

---

Quick Start

1. Setup AlloyDB

bash setup_alloydb.sh YOUR_PROJECT_ID

2. Initialize Database

# Apply schema
psql -h YOUR_ALLOYDB_IP -U postgres -d pragati -f db/schema.sql

# Seed with India health data
pip install -r requirements.txt
python db/seed_data.py

3. Run Locally

cp .env.example .env
# Edit .env with your AlloyDB IP and GCP project
uvicorn api.main:app --reload --port 8080

4. Deploy to Cloud Run

bash deploy.sh YOUR_PROJECT_ID

---

API Endpoints

Endpoint	Description
GET /	Web UI dashboard
POST /query	Natural language health query
GET /tools	List all self-assembled MCP tools
GET /boot-log	See the self-assembly boot sequence
GET /health	Health check
GET /stats	Tool usage statistics

---

Sample Questions

• "What is the infant mortality rate in Bihar?"
• "Compare immunization coverage across all states"
• "Show malaria hotspots in 2024"
• "How many PHCs are there in Rajasthan?"
• "What are TB detection rate trends in Maharashtra?"
• "Dengue cases in Tamil Nadu"
• "Facility summary by state"

---

Data

PRAGATI uses HMIS-style India public health data across:

• 10 states x 6 districts each
• 15 health indicators (IMR, MMR, immunization, etc.) x 4 years
• 600+ health facilities (PHC, CHC, hospitals)
• Disease surveillance for 8 diseases x 3 years
• 7,894 total rows across 3 tables

---

Team

Nipun Sujesh | Google Gen AI Academy APAC Cohort 1 | April 2026

Built with Google ADK, MCP Toolbox, AlloyDB, Vertex AI, and Cloud Run.