B1
Track B — Multi-Agent Orchestration & Decision Support

Confidential Document Q&A Workbench

Systems where multiple specialized agents collaborate, hand off tasks, and reach a combined decision.

Background

Organizations that handle sensitive information — refineries, PSUs, defence-adjacent industries — cannot send confidential documents to external cloud AI APIs, but still need AI-assisted document workflows.

Objective

Develop a self-contained, on-premise "AI workbench" where users interact with multiple specialized agents — summarizer, Q&A, comparator — over their own uploaded documents, using only open-weight models.

Key Features

1. Document upload into a local knowledge base (PDF/DOCX) 2. Summarizer agent 3. Q&A agent (RAG over uploaded documents) 4. Comparator agent (structured diff between two documents/sections) 5. Orchestrator that routes requests to the right specialist agent 6. Fully local/offline execution (no external LLM API calls)

Expected Solution

A multi-agent orchestration layer that decides which specialist(s) to invoke Three working specialist agents (summarizer, Q&A, comparator) A local vector store and locally-hosted open-weight LLM

Suggested Technology Stack

Model hosting: Ollama (Llama 3.1 8B, Mistral 7B, Phi-3) Orchestration: LangGraph Vector store: ChromaDB (local) Frontend: Streamlit

Expected Outcomes

A reusable in-house agentic AI pattern for confidentiality-sensitive organizations Reduced dependence on external cloud AI for internal document work Demonstration that useful agentic systems don't require paid APIs

Possible Use Cases

Internal policy document Q&A for employees Comparing two versions of a contract or SOP Quick summarization of long internal reports

Evaluation Parameters

Correctness of orchestration (right agent for the right request) Quality of each specialist agent's output Genuine local/offline operation Explainability of which agent handled what