Project Vault
Air-gapped on-premise agentic AI workbench using open-weight multimodal LLMs for confidential industrial work
- Year
- 2026
- Role
- Lead Systems & Agentic AI Architect
- Context
- Smart India Hackathon 2026 · MRPL (Problem ID 26117)
- Status
- Completed
Overview
Project Vault is a sovereign, fully self-hosted agentic AI workbench designed for public sector undertakings (PSUs), refineries, and defense organizations where sensitive operational data cannot be transmitted to external cloud providers.
Engineered for Smart India Hackathon (SIH) 2026 under problem statement 26117 for Mangalore Refinery and Petrochemicals Limited (MRPL), it combines autonomous multi-step agent reasoning, document intelligence, local RAG, and sandboxed code execution.
An integrated network monitor and append-only audit trail continuously verify that zero external API requests or outbound connections are made during operation.
Key Highlights
What makes it standout
- Autonomous agent orchestrator featuring TaskRouter, TaskPlanner, AgentExecutor, and verification loops
- Local multi-model inference powered by llama.cpp running open-weight models (Qwen3-4B, Gemma-3-4B)
- Hybrid RAG knowledge pipeline leveraging local Qdrant vector database and grounded citation tracing across PDFs & DOCX
- Ephemeral Docker code sandbox (python:3.11-alpine, --network none) with real-time xterm.js streaming terminal
- Network monitor and tamper-evident audit logging providing cryptographic sovereignty proof of air-gapping
System Design
Architecture & Pipeline
