Zero-leak
by architecture.
A privacy policy is a promise. An architecture is a constraint. Here is exactly where your data goes in an Divishta pipeline, and everywhere it doesn't.
Every hop, accounted for.
| Stage | Where it runs | What protects it | What is stored |
|---|---|---|---|
| 1 · Ingestion | Our private server, or your VPC / on-prem | TLS 1.2+ in transit; scoped, revocable connector credentials | Your documents, encrypted at rest |
| 2 · Indexing | Same environment as ingestion | Per-tenant indexes; embeddings computed locally | Chunks + vectors, tagged with access rules |
| 3 · Retrieval | Same environment | Permission check per user before the model sees text | Nothing new |
| 4 · Inference | Self-hosted open-weight model (llama.cpp) | Bound to a private interface; no external network egress | Nothing: prompts are not logged by default |
| 5 · Actions | Your systems, via approved tools only | Allow-listed tools, step and cost budgets, human approval for consequential actions | Audit event (who, what, when; no content) |
| 6 · Deletion | All of the above | On-request erasure of documents, vectors and backups | Deletion record |
What we do, and what we never do.
No third-party AI by default
The default pipeline makes no calls to OpenAI, Anthropic, Google or any external model API. Using one for a specific task needs your written opt-in.
No prompt or answer logging
Inference servers run with request logging off. We keep operational metrics (latency, token counts), never content.
Never used for training
Your data is retrieved at question time. It is never used to train or fine-tune shared models.
Isolation in code, not in prompts
Tenant and user permissions are enforced in the retrieval layer. A cleverly worded question can't reach another tenant's data.
Your cloud, if you prefer
Every component ships as containers that run in your VPC or on-premise, with local models and no mandatory external calls.
Breach response
A documented incident process aligned with India's DPDP Rules, including prompt notification so you can meet your own reporting duties.
Under DPDP, the data fiduciary remains responsible even when a processor does the work. That includes your AI vendors. Self-hosted inference shrinks the list of processors you have to vouch for.
What we offer every client
Data processing agreement
Data map of every system the agent touches
Retention & erasure schedule
Audit log export on request
What we don't claim (yet).
We are a young company. We do not currently hold ISO 27001 or SOC 2 certification, and we won't imply that we do. What we can show you today is the architecture itself: you can inspect the deployment, run it inside your own network, and verify that no external AI endpoint is called.
"Zero leak" means no third party in the data path by design. No system is immune to every failure. That's why we pair it with least-privilege access, encryption, audit trails and a breach process. Independent audits are on our roadmap.
Even this website practises what it preaches: no trackers, no analytics, no external fonts and no CDNs. Open your browser's network tab and check.