2025 · AI agent layer
BharatBiz AI
Multi-agent assistant for Indian small businesses
BharatBiz AI was a hackathon bet: that a single agent — not a suite of apps — could shoulder customer support and inventory questions for a small business.
- Year
- 2025
- Role
- AI agent layer
- Status
- Open source
- Repository
- github.com/Ghoulayush/BharatBiz-AI
01Overview
A multi-agent system built for the Code for Bharat — Season 2 hackathon that automates customer support and inventory workflows for small businesses. I built the AI agent layer: a LangChain ReAct agent on Cerebras that decides between an EMI calculator, an inventory-advice tool, and a Pinecone-backed retriever over government schemes.
02The problem
Indian small businesses run on manual workflows and have little access to affordable tooling for customer service, inventory tracking, and financial planning.
03Goals
- 01Automate routine customer-support and inventory workflows.
- 02Let one agent pick the right tool per query instead of forcing separate flows.
- 03Surface relevant government schemes to small-business owners.
- 04Ship a working entry for Code for Bharat — Season 2 within hackathon constraints.
04Key features
- Built a LangChain ReAct agent running on Cerebras
- Agent routes between an EMI calculator, inventory advice, and a Pinecone-backed schemes retriever
- Entry for the Code for Bharat — Season 2 hackathon
05Architecture overview
01The agent layer is a LangChain ReAct agent running on Cerebras. ReAct gives it a reasoning step before acting, so each incoming query is turned into a tool call.
02The agent chooses between three tools: an EMI calculator for loan questions, an inventory-advice tool for stock decisions, and a Pinecone-backed retriever that searches embedded government schemes.
03A FastAPI layer exposes the agent to clients, while the React and TypeScript frontend provides the interface, with Firebase handling auth and data.
06Challenges faced
- Getting tool routing right: the agent must pick the correct tool from a loosely worded question.
- Retrieving schemes that are actually relevant to a given business context.
- Keeping the whole flow coherent under hackathon time pressure.
07Engineering decisions
ReAct over a hardcoded decision tree
Giving the model a reasoning step before each tool call handles the messy phrasing real users bring.
Cerebras for inference
Low-latency inference keeps the agent's reasoning-and-act loop responsive enough for interactive support.
Pinecone for scheme retrieval
Vector search over government schemes matches the query's intent rather than exact keywords.
08Lessons learned
- An agent is only as good as its tools — the EMI calculator and retriever define what it can actually do.
- Fast inference changes the feel of agentic loops more than any single prompt tweak.
- Hackathon scope is an exercise in deciding which workflows deserve a tool at all.
09Future improvements
- 01Add memory so returning customers do not repeat context.
- 02Expand the tool set beyond support and inventory.
- 03Evaluate routing accuracy with a small labeled dataset.
Media gallery
Placeholder — swap in real captures