SIH26117: Sovereign On-Premise Agentic AI Workbench using Open-Weight Multimodal LLMs for — SIH 2026 | Wynit

SIH26117: Sovereign On-Premise Agentic AI Workbench using Open-Weight Multimodal LLMs for — SIH 2026 | Wynit

What they'll ask you

12

Generated from the brief's own wording and the competition's published rules, never from a guess about what this ministry prefers.

Specific hardware models or GPU VRAM limits available at the venue?
The brief never answers this, so a panel will. Whatever you decide, say it the same way twice.
Exact open-weight models preferred by the organization?
The brief never answers this, so a panel will. Whatever you decide, say it the same way twice.
Has any part of this been shown at a previous event, hackathon or college project?
The guidelines are explicit: your solution must not have appeared in any previous event or programme, of any sort. A recycled project is what a team under time pressure reaches for.

9 morethe ones a panel gets to if there is time

The brief asks for local backend and routing. How would you build that?
Decoded from SIH26117 itself: A backend that supports multiple open weight models and routes tasks to the correct model based on requirements. The brief asks for it by name.
The brief asks for agentic task engine. How would you build that?
Decoded from SIH26117 itself: An engine that plans multi-step work, executes local tools, and iterates on tasks. The brief asks for it by name.
The brief asks for multimodal and ocr processor. How would you build that?
Decoded from SIH26117 itself: A component using on-device OCR and vision models to process scanned PDFs, drawings, and images. The brief asks for it by name.
The brief asks for local knowledge base connector. How would you build that?
Decoded from SIH26117 itself: A connector that grounds responses in local manuals, SOPs, and past correspondence. The brief asks for it by name.
The brief asks for deliverable generator and sandbox. How would you build that?
Decoded from SIH26117 itself: A system to run code in a sandbox and output real files like Word, Excel, and presentations. The brief asks for it by name.
Why not use what already exists? Name the closest thing to this that is already running.
A team that has not named the alternative themselves is answering this for the first time in the room.
Which single thing will you demonstrate end to end, start to finish, with nothing skipped?
Ours, not a rule: a narrow thing that fully works survives questioning better than a broad thing that half works. If nobody on the team can name it, that is the finding.
Show me this working: the jury will check whether the system operates entirely offline using visible network monitors or logs to prove zero external calls.
This is the evaluator read for your problem statement, decoded from the brief's own wording.
Show me this working: they will test model auto-selection across different tasks, end-to-end agentic workflows, document processing, and sandbox code execution.
This is the evaluator read for your problem statement, decoded from the brief's own wording.

The government's original description (462 words)

• Background Refineries, PSUs, defence-linked manufacturing units and government offices generate a lot of routine but sensitive knowledge work. Approval notes, board presentations, engineering calculations, code for internal tools, review of scanned drawings and inspection reports. None of this can go through cloud AI assistants like Claude or Codex because the underlying data is confidential: Piping & Instrument Diagrams, financials, vendor negotiations, unreleased designs, internal correspondence, confidential business strategies etc. Company policy keeps this data on premises, so people either do the work manually resulting in productivity gain, or they quietly paste confidential material into public tools anyway. Open weight large reasoning models have reached a point where a genuinely useful assistant built on them is realistic. But nothing deployable exists today that industrial users can actually work with the way they use Claude or Codex. • Description The idea is a self-hosted, air gapped AI workbench running entirely on the organization's own GPU server. Nothing leaves the premises. The backend should not be locked to one model. It needs to support multiple open weight models at once and automatically pick the right one for a given task based on what that task needs, a coding request handled differently from a document summary request. New open weight models should be addable later without redesigning the system, since this space is moving fast. The assistant also needs to actually act like an agent. Plan out multi step work, call local tools such as file read and write, code execution in a sandbox, spreadsheet work, internal document search, and iterate on a task instead of answering once and stopping. It needs to handle more than text too: scanned PDFs, handwritten notes, engineering drawings, photographs, read through on device OCR and vision models. Output should be real deliverables, approval notes, PPT/Word/Excel files, working code, calculations with steps shown, not just chat replies. And it needs to ground itself in the organization's own manuals, SOPs and past correspondence through a local knowledge base connector, again with nothing going external. • Expected Solution A working local deployment, demonstrable on a single workstation or server with a mid range GPU (use a smaller open weight model if 120B class hardware isn't available at the venue), that shows model auto selection across at least two different task types. An agentic task carried through end to end, for example reading a scanned inspection report, pulling out key findings and drafting an approval note as a Word file. A coding task run and verified in a sandbox. A multimodal task involving image or scanned document understanding. The system should also show, through logs or a visible network monitor, that no external calls are made at any point. That's the actual proof of the sovereign claim, not just a statement of it.