Runs where your data already is
Deployed on your hardware, in your VPC, or fully air-gapped. The model comes to the data rather than the data going to the model, so no personal data crosses a boundary to be processed.
Overview
A small language model built in-house and deployed inside your estate, so the personal data it reasons over never leaves the perimeter — and no third-party model provider ever enters your processing chain.
Why build our own
A hosted model is a processor you did not vet
Every other control on this platform exists to keep personal data accounted for. Routing that same data through someone else's endpoint to summarise it would undo the lot.
- A prompt is a disclosure
- The moment a customer record, a claim file or a KYC document is pasted into a hosted model, it has been transferred to a processor — usually one outside India, usually under terms you did not write. DPDP does not have an exception for convenience.
- You cannot audit what you cannot see
- A third-party endpoint returns an answer and nothing else. There is no record of what the model was trained on, what it retained, or which version produced the output you are now relying on in a decision.
- Retention you do not control
- Provider retention windows, abuse-monitoring copies and training opt-outs are contract terms, not controls. A model running on your own hardware has no such surface to negotiate.
What it does
Sovereign by construction, not by contract
The guarantees come from where the model runs, so they hold without depending on a provider honouring a term you cannot verify.
No egress, no third-party processor
Nothing leaves the perimeter — not prompts, not documents, not embeddings. There is no model vendor in your processing chain, so there is no vendor to add to your records or your DPA schedule.
Tuned on your corpus, not the public web
Adapted to your policies, product names and document formats, so it answers in your terms. The tuning data stays yours and is never pooled to improve anyone else's model.
Every inference is on the record
Prompt, model version, retrieved context and output land in the same append-only trail the rest of the platform writes to, so an automated output can be reconstructed months later.
Governed like any other model
It appears in the model register with its owner, purpose and training-data provenance, and is assessed against the EU AI Act and DPDP Sec. 10 in the same engine as every other system you run.
Small enough to be practical
Sized to run on commodity GPU hardware rather than a datacentre, which is what makes on-premises deployment a real option instead of a procurement exercise.
Deployment
Three ways to run it
Pick the one your obligations require. The model is the same in all three; only the boundary moves.
- 01
On-premises
Your servers, your network, your access control. The usual choice for banks and government bodies with data-localisation obligations.
- 02
Private cloud / VPC
A dedicated tenancy in your own cloud account and region. Nothing is shared with another customer and nothing routes through us.
- 03
Air-gapped
No outbound network at all. Model and weights are delivered and updated out of band, for classified and critical-infrastructure environments.
Where it sits
The SLM is one of seven Data Security categories, and it is governed by the same programme that governs the detection models — registered, assessed, and evidenced alongside everything else that touches personal data.
All of Data SecurityRun it against your own documents
A walkthrough deploys the model in a sandbox that mirrors your constraints — on-prem, VPC or air-gapped — and puts your material through it.