Janmitra: The AI Agent

Inspiration

We were inspired by the countless civic complaints that flood social media every day—reports of water leaks, garbage pile-ups, and broken roads that often go unheard. These posts represent a critical failure in citizen-government communication. We saw an opportunity to leverage AI not just as a conversational tool, but as an active facilitator of democracy. Our inspiration was to build a digital bridge, an agent that could autonomously listen, understand, and act on behalf of citizens, transforming social media frustration into tangible government action.

What it does

Janmitra is an autonomous AI agent that automates civic grievance redressal. It continuously monitors social media platforms for public complaints, intelligently classifies them based on issue type (water, roads, garbage) and location using natural language processing, and automatically files them into the appropriate government's grievance portal. It then provides the citizen with a tracking ID, closing the feedback loop. Essentially, it turns a tweet into a ticket, ensuring no plea for help is lost in the digital void.

How we built it

Janmitra is built on a robust, open-source stack centered around IBM's technologies, following a precise agentic workflow.

Core Orchestration: The IBM Agent Development Kit (ADK) serves as the agent's brain, managing its state and the reasoning loop.

Reasoning Engine: The IBM Granite-13B-Instruct model, run locally via Ollama, provides the NLP muscle for understanding vernacular language and classifying complaints.

Tools & Capabilities: We equipped the ADK agent with a suite of custom tools:

A TwitterScraperTool to fetch mock tweets from a curated dataset.

An NLPClassifierTool that uses the Granite model to analyze text.

A GovernmentAPITool that simulates submitting a ticket to a mock government API.

Memory & Dashboard: All actions are logged to a SQLite database and visualized in real-time on a Streamlit dashboard, providing a window into the agent's operations.

Challenges we ran into

The Vernacular Wall: Processing mixed-language queries (Hinglish, Tanglish) was a significant initial hurdle. The model often struggled to correctly identify locations and issues from colloquial language.

Agentic Orchestration: Designing a reliable flow within the ADK, where the agent correctly sequences tool calls (e.g., first classify, then submit) and handles errors gracefully, required extensive iterative testing.

Resource Constraints: Running a large language model (LLM) locally on consumer hardware forced us to optimize for efficiency, balancing response quality with inference time and memory usage argmax settings P(Performance∣Latency,Memory)

Accomplishments that we're proud of

Creating a True End-to-End Agent: We built a fully functional system that goes from a raw social media post to a confirmed government ticket without any human intervention. This complete automation is our biggest achievement.

Mastering the IBM Stack: Successfully integrating the IBM ADK with the IBM Granite model to create a cohesive and powerful application, exactly as the track intended.

Solving a Real Problem: Building a project that has a direct, positive social impact and addresses a daily pain point for millions of citizens makes us immensely proud.

What we learned

This project was a deep dive into the practicalities of modern AI. We moved beyond theory and learned:

The paradigm of Agentic AI: How to architect systems where LLMs act as reasoning engines orchestrating tools.

The power of open-source: How to leverage and integrate powerful open-source models (IBM Granite) and frameworks (ADK) to build cost-free, sophisticated applications.

Full-stack AI development: Integrating various components—LLMs, databases, APIs, and front-end dashboards—into a single, seamless pipeline.

Prompt engineering for robustness: Crafting precise context and instructions for the LLM to reliably perform complex, multi-step tasks.

What's next for Janmitra: The AI Agent

Real API Integrations: Partnering with government bodies to integrate with live APIs like the Integrated Grievance Redressal System (IGRS).

Multimodal Intelligence: Incorporating image recognition to allow citizens to submit photos of problems (e.g., potholes, water leaks) for automatic analysis.

Advanced NLP for Regional Languages: Fine-tuning the model on a wider corpus of Indian languages to drastically improve vernacular understanding.

Deployment and Scaling: Exploring cloud deployment to run this as a live public service, truly democratizing access to governance.

Janmitra is a proof-of-concept for a more responsive government, and our journey to make it a reality has just begun.