Sovereign AI for defence, DPSUs and PSUs, on your own servers
BiltIQ sovereign AI for defence runs Agent OS and AI Factory on a DPSU's or PSU's own servers. Agents read tenders and contracts, check them against your rules and cite the page. Nemotron is the default model, outbound calls are blocked by default, and high-consequence actions wait for a named approver. Work starts with an NDA and a paid pilot.
We'll walk you through it on sample data, never on real defence documents.
Scrutiny by hand, and every model's origin questioned
| Delivered late | Procurement contracts are often delivered late. |
|---|---|
| Checked by hand | Tenders and contracts are read page by page, often over months. |
| Model origin | Defence and PSU buyers check each model's origin under Public Procurement No. 1. |
| Data cannot leave | Some files may not go to a cloud API at all. |
Open-weight AI models for defence: Nemotron by default
Model-agnostic: open-weight models chosen per buyer. NVIDIA Nemotron by default for defence and PSU; Qwen3.8-27B for reasoning where allowed.
For defence and PSU work we propose NVIDIA Nemotron models by default. Most Nemotron models are released under the NVIDIA Open Model License; some embedding, reranking and parsing models carry other stated licences as well, such as the Llama 3.2 Community License or OpenMDW. These are open-weight licences, not OSI open-source licences; we say so up front. You choose the models; we configure them.
On your own servers, with outbound calls blocked
Agent OS and AI Factory run on your GPU server, inside your walls. Outbound calls are blocked by default (platform gate), and an unconfigured install blocks all outbound calls. In on-prem-required mode, no outbound host can be allowed at all. Patent application 202631094208 covers the gate; it is not yet granted. Serving runs a single node per model today; multi-node serving is written, not yet in use.
AI for tender and contract scrutiny: cited findings, a person decides
Document in.
Your file is loaded on your server. With the gate on, nothing goes to an outside API.
Agents check.
Your SI agents test clauses, amounts and dates against your checklist.
Cited findings.
Each finding shows the page and the text it rests on.
Approval inbox.
A named reviewer accepts or rejects each finding. Irreversible and high-consequence actions pause for human approval; if no approver is available, they are blocked.
97 agents built; 11 pass all 7 of our quality checks.
A paid pilot under NDA, on your own past files
- Terms: NDA first. ₹5–8 L, ex-GST, 4–6 weeks, credited against the licence.
- Where: on your GPU server or a BiltIQ loaner; your documents stay on site.
- Measure: past files with known outcomes: caught, missed, time per file. You get a findings report against your manual results, then make the licence decision.
- Pilot lead: Harish Kumar, CEO (18 yrs).
This is a plan, not a measured result; we quote no speed-up.
Own it after the pilot: site licence ₹50–90 L one-time + managed service ₹16–18 L a year; hardware at cost; ex-GST. The managed service covers updates, monitoring with no document access, a quarterly review and agent tuning; business hours IST support, 4-hour site-down response.
Buying from BiltIQ
BiltIQ AI is Aarna Tech Consultants Pvt. Ltd., a DPIIT-recognised startup (DIPP239966) based in Jamshedpur. We are ISO/IEC 27001:2022 and ISO 9001:2015 certified and an NVIDIA Inception Partner. We work to your procurement rules; the route is agreed on the briefing call.
Proof so far
Our public proof today is the NHA PM-JAY AI Challenge 2026: won (hackathon), claim and bill verification with Agent OS, IISc Bangalore finale, 8 May 2026. Every walk-through we give uses sample data.
Key facts: Defence, DPSU and PSU
- Model-agnostic: open-weight models chosen per buyer. NVIDIA Nemotron by default for defence and PSU; Qwen3.8-27B for reasoning where allowed.BiltIQ code review, Oct 2026
- Outbound calls are blocked by default; an unconfigured install blocks all outbound calls; in on-prem-required mode no outbound host can be allowed.BiltIQ code review, Oct 2026
- Irreversible and high-consequence actions pause for human approval; with no approver, they are blocked.BiltIQ code review, Oct 2026
- Paid pilot: NDA first, ₹5–8 L, 4–6 weeks, credited against the licence; documents stay on site on your server or a BiltIQ loaner.BiltIQ price list, Oct 2026
- Patent application 202631094208 (fail-closed data boundary), filed Aug 2026, not granted.Indian Patent Office filing
- Winner, NHA PM-JAY AI Challenge 2026 (hackathon), IISc Bangalore finale, 8 May 2026.NHA declaration, May 2026
FAQ
Can AI for defence run fully on our own servers?
Yes. Agent OS and AI Factory run on your GPU server, inside your walls. Outbound calls are blocked by default by the platform gate, and an unconfigured install blocks all outbound calls. In on-prem-required mode, no outbound host can be allowed. Your SI's models, agents and documents stay on your hardware; we monitor health remotely only with your permission and with no document access.
Which AI models do you use for defence, and where do they come from?
BiltIQ is model-agnostic: open-weight models chosen per buyer. NVIDIA Nemotron by default for defence and PSU; Qwen3.8-27B for reasoning where allowed. Defence buyers check each model's origin under Public Procurement No. 1, which is why Nemotron is the default there. These models carry open-weight licences, not OSI open-source licences.
How does AI tender scrutiny work for a DPSU?
Your file is loaded on your own server. Your SI agents test clauses, amounts and dates against your checklist, and each finding cites the page and text it rests on. A named reviewer accepts or rejects every finding in the approval inbox. We prove it on your own past files in a 4–6 week paid pilot, against your manual results.
What does an AI pilot for a PSU cost?
The paid pilot is ₹5–8 L, ex-GST, for 4–6 weeks on your own past documents, and the full fee is credited against the licence. It starts with an NDA and runs on your GPU server or a BiltIQ loaner, so your documents stay on site. After it, a site licence costs ₹50–90 L one-time plus ₹16–18 L a year.
Can an AI agent act without a person approving it?
Not for the steps that matter. In Agent OS, irreversible and high-consequence actions pause for human approval, verified in our code. Findings land in an approval inbox, where a named reviewer accepts or rejects each one. If no approver is available, the action is blocked rather than run. Reading and checking run on their own; consequential steps wait.
Request an NDA-first briefing
We'll walk you through it on sample data. Or write to [email protected], or call or WhatsApp +91 89868 60088 (messages go straight to our CRM).