Sovereign On-Premise Agentic AI Workbench using Open-Weight Multimodal LLMs for Confidential Industrial Work
Mangalore Refinery and Petrochemicals Limited (MRPL) · Smart Automation · Software
The best-framed software problem in this block and the network-cable demo is unbeatable, but the scope is six systems in one — pick the agent loop and real file deliverables as your depth, and be honest about what the router does.
Data: Open-source models and publicly available document samples (sample scanned PDFs, sample P&IDs from open datasets)
What it actually is
Refineries and defence-linked units do a lot of routine knowledge work — approval notes, engineering calculations, reviewing scanned drawings — that cannot go near a cloud AI tool because the material is confidential. So people either do it manually or quietly paste sensitive documents into public chatbots anyway. The ask is a self-hosted assistant running entirely on the organisation's own hardware, using open-weight models, that actually works the way people use commercial AI tools.
What to build
An air-gapped workbench running open-weight models locally, with a routing layer that holds several models at once and picks the appropriate one per task — a coding request routed differently from a document summary — and accepts new models without redesign; an agent loop that plans multi-step work and calls local tools including file read and write, sandboxed code execution, spreadsheet manipulation and internal document search, iterating rather than answering once; multimodal ingestion of scanned PDFs, handwritten notes, engineering drawings and photographs through on-device OCR and vision models; retrieval grounding against the organisation's own manuals and documents; and output as real deliverables — Word, Excel and PowerPoint files, working code, calculations with steps shown — rather than chat replies.
Smallest thing that wins the room
With the network physically disconnected, hand it a scanned P&ID and an internal standard, and watch it read the drawing, ground its reasoning in the retrieved document, run a calculation showing its steps, and produce a formatted approval note as a downloadable Word file.
How crowded this one gets
A guess, projected from the 2025 statements — the last year where both the submission counts and the winners were published.
Quieter than 73% of the 226 · #61 of 226 by expected field
A normal-sized field. Your idea has to be good, not miraculous.
Why: company-sponsored statements drew the smallest fields of all.
This is a guess, not a fact
Nobody has published 2026’s numbers yet. This is an analysed estimate from last year’s pattern, so please do not take it as the truth — check the live counter on the SIH portal before you decide anything. The range covers the middle half of likely outcomes, so one statement in two lands outside it. Entry closes at 500 ideas per statement, so no range goes past that — a statement that reaches the cap fills and shuts rather than drawing an unlimited crowd. The model reads only three things a team can see before choosing — software or hardware, the theme, and what kind of body posted it — and those explain about a quarter of the variation in last year’s field sizes (R² 0.25 on held-out statements). Trust the band more than the number, and the ordering more than either. It cannot see how good your idea is, which is the part that actually decides it.
The scores
The number is the shorthand. The line under it is the reason.
Acceptance potential
4/5The problem is real, current and well-articulated, the sovereignty framing lands hard with a PSU judge, and the field will be thinned by the GPU requirement — but the scope is enormous and a team that builds a local chatbot without the agent loop or the file deliverables has answered a fraction of it.
Feasibility
3/5Open-weight models, local inference servers, agent frameworks and document generation libraries all exist and compose, but the description asks for model routing, agentic tool use, sandboxed execution, multimodal OCR and RAG together — and running several open-weight models concurrently needs GPU capacity most teams do not have.
Innovation scope
4/5The routing policy, the agent architecture and how you make an air-gapped system genuinely useful rather than merely offline are all undefined, and the description explicitly leaves the model mix and extensibility mechanism to you.
Clarity
4/5The description is unusually concrete about what must exist — multi-model routing, extensibility, planning and iteration, named local tools, multimodal inputs, real file deliverables, and grounding in organisational documents — even though it sets no performance targets.
Effort
MassiveModel serving and routing, an agent loop with tool calling, a code sandbox, a multimodal document pipeline, a retrieval layer and Office-format generation is essentially rebuilding a commercial AI coding assistant locally.
Demo-ability
EasyPulling the network cable and then watching it produce a real Word document from a scanned drawing is a dramatic, self-evident demonstration that needs no explanation.
In its favour
- Green flag: Disconnecting the network during the demo is a single physical action that proves the entire sovereignty claim more convincingly than any slide could
- Green flag: The description names its own benchmark — users should work with it the way they use commercial assistants — which gives you a clear and honest bar to aim at
- Green flag: Producing an actual Word or Excel file rather than chat text is a modest engineering addition with outsized perceived value for an industrial user
- Green flag: The confidentiality problem is genuine and immediately recognisable to any PSU judge, so no persuasion is needed on why it matters
Against it
- Red flag: Serving several open-weight models concurrently needs real GPU memory, and without access to a capable machine you will demo one small model and quietly drop the routing that the description asks for
- Red flag: The scope spans routing, agents, sandboxing, multimodal ingestion, retrieval and document generation — six substantial systems, and building all six shallowly reads worse than building three properly
- Red flag: Model routing is easy to assert and hard to justify, so be ready to show that your router actually improves outcomes rather than adding a layer that looks sophisticated
- Red flag: Engineering drawings and P&IDs are dense symbolic documents that general vision models read poorly, so the multimodal claim will fail on exactly the input the description highlights
What you will be writing
- vLLM or Ollama local open-weight serving
- Qwen / Llama / Mistral multimodal open-weight models
- Agent loop with tool calling (LangGraph or custom)
- Docker-sandboxed code execution
- On-device OCR + vision (PaddleOCR, docling)
- python-docx / openpyxl / python-pptx deliverable generation
- Sovereign AI infrastructure
- Agentic systems
- Industrial knowledge work
Prior art to read before you start
air-gapped local LLM deployment · agentic multi-step task execution · document-grounded enterprise assistant
Analysed by Claude Opus. Every score above is a judgment call with its reasoning attached — kindly cross-check this against the official statement on the SIH portal before your team commits to it.