Private · Sovereign AI
Private & Sovereign AI on your hardware.
Some data can't leave the building. Banks, healthcare, defence, government — and increasingly mid-market firms with paying customers in Europe — need the model to run inside their boundary. We architect the cluster, deploy the model, and tune the inference path without compromising the audit trail.
- Open-weight families — Llama, Mistral, Qwen, Phi — chosen by capability + licence.
- vLLM, Ollama, NVIDIA Inference Microservices, custom — picked by throughput need.
- GPU sizing math from first principles — no vendor influence.
- Audit trail and model-card baked in to satisfy regulator review.
“Sovereign AI isn't about distrust of foreign models. It's about owning the failure mode when the API rate-limits at 3am.”
§ 02 / Authority
From the roomsI've been in — and the feeds that came out of them.
Live shorts. LinkedIn originals. Unedited.
Read all 11 field notes →
PMI Gujarat · Keynote · Dec 2024
I had the incredible opportunity to be part of the PMI Gujarat Chapter — NEXTGEN Project Management, elevating our world through emerging technologies.
Live · this quarter
48,000+
LinkedIn impressions
- Field notes11
- Industries20
- Years shipping15+
Karnavati · AI Conclave 2025
Unlocking the Power of AI: it's all about the right prompt. At Karnavati University, I had the privilege of sharing a key insight…
Rajkot · Keynote · 2025
What happens when AI meets India's social and rural future? You get a room full of students, academics, and leaders in Rajkot asking the questions that actually matter.
PM Forum · Off-script · 2025
I walked into a room of 45+ project managers. No slides. No deck. No script.
- PMI Gujarat
- Karnavati University
- Atmiya AI Summit
- HP WeRise · Ahmedabad
- Project Managers' Forum
- NIFT Gandhinagar
- MeitY · Digital India
- UIT Karnavati
- Ahmedabad founder circle
- PMI Gujarat
- Karnavati University
- Atmiya AI Summit
- HP WeRise · Ahmedabad
- Project Managers' Forum
- NIFT Gandhinagar
- MeitY · Digital India
- UIT Karnavati
- Ahmedabad founder circle
What this means in practice
We treat sovereignty as a system property, not a buzzword. The model has to run, the prompts have to log, the embeddings have to live somewhere — all inside the boundary, all auditable.
- 01
Architecture document — cluster topology, model selection, throughput math, failure modes.
- 02
Reference deployment — vLLM / NIM / Ollama with monitoring, autoscaling, logging.
- 03
Audit log + model-card pack ready for regulator + customer review.
- 04
Operations runbook — patching cadence, rollback procedure, capacity planning.
Engagement shapes
2 ways to engage on Private & Sovereign AI. Pick the closest fit; we calibrate scope on the diagnostic call.
01
2 wk
Sovereign-AI feasibility
Read your data, regulatory, and capability needs. Output: a written go/no-go on private deployment + the rough cost shape if you go.
02
8–16 wk
Embedded build
Architecture + deploy + monitoring + runbook. Embedded with your platform team. Hand-off includes a 30-day on-call window during ramp.
For
Who this fits
- Banks, NBFCs, insurers under RBI / IRDAI scrutiny.
- Healthcare orgs handling PHI under HIPAA-equivalent / DPDP rules.
- EU customers under GDPR + EU AI Act with hard data-residency clauses.
- Defence / government agencies with classified or strategic data.
Not for
Honest gates
- Teams whose only reason for sovereign is FUD about foreign clouds — there are cheaper hybrid postures.
- Anyone unwilling to budget for GPU capex / opex realistically.
FAQ
About this service.
Which open-weight models do you recommend?
Llama 3.x family for general-purpose, Qwen 2.5 / 3 for strong multilingual + code, Mistral Large for European licensing certainty, Phi-3 for edge. The right choice depends on your capability needs + licence + GPU budget — we score against 4–6 candidate models on your real workload, not on benchmarks.GPU sizing — how much do we actually need?
First-principles math per project — token budget × concurrency × latency target. As a rough guide: 8B-class model in fp16 fits on a single A10G; 70B-class in fp8 needs 2× A100 80 GB; concurrency scales linearly with VRAM headroom. We size with explicit failure-mode planning (what happens at 2× peak load, what happens during model swap).RBI / IRDAI compliance support?
Yes — we've delivered RBI-aligned voice + AI deployments. The architecture: every model decision audit-logged, every prompt + response logged with sequence ids, model-card pack with training-data lineage, change-control process aligned to the regulator's IT outsourcing framework. Maps to the same governance system as ISO 42001.How do you handle model updates?
Treated like a database migration. Canary deploy (5% traffic), eval suite runs against the new model, regression alerts halt the rollout, rollback plan documented before the deploy. Open-weight models are easier than API-based — you control the timing.Cost compared to using a cloud LLM API?
Break-even is around 5–10 M tokens / day. Below that, API is usually cheaper; above that, your own GPUs win on per-token cost (and always win on data residency). We model this for your specific workload during the feasibility phase, not on industry averages.Network architecture — air-gapped or cloud-hybrid?
Both shapes ship. Air-gapped: cluster sits inside your secure VPC, model registry lives on internal storage, updates are blue-green via signed manifests. Cloud-hybrid: control plane in your cloud, data plane in a regional VPC that respects residency. We've shipped both for BFSI and for government.
Related services
You may also need these.
01
AI Governance
ISO 42001 · NIST RMF · EU AI Act · DPDP — operationalised.
Read the page →02
AI-as-a-Service
Strategy → architecture → ship — one accountable engineer.
Read the page →03
LLMOps
Evals, observability, and deploys that survive Monday morning.
Read the page →
Voices from the rooms
48,000+ impressions · 11 notes · 9 venues
PMI Gujarat · Keynote · Dec 2024 Karnavati · AI Conclave 2025 Rajkot · Keynote · 2025 PM Forum · Off-script · 2025
Next step
Tell me what you’re trying to ship.
One brief, one inbox, one reply within a business day. Either a calendar slot, a referral, or a straight no — we don’t bench-fit work that isn’t a fit.