Megh Vyas
Backend · Infra · SRE
← All work

Zero-trust AI

ZTA-AI

Ask your company's data anything, without ever letting the language model see the data itself.

Year
2025
Role
Systems & security
Built with
FastAPI, Postgres, Redis, RBAC / ABAC, Docker

<0ms

p95 pipeline latency

0

raw data to the model

0

compliance frameworks

The threat model

Enterprises want to ask their internal data questions in plain language. What they cannot do is hand raw records, schemas, or credentials to a model that might memorize or hallucinate them. The usual RAG shortcut quietly stuffs all of that into the prompt and calls it a day.

The controls

  1. 01

    Separated data access, business logic, and language generation completely. A deterministic compiler makes every decision; the model only renders an approved answer.

  2. 02

    Modeled all data as immutable, versioned claims carrying provenance, sensitivity classification, and compliance tags.

  3. 03

    Gated access with department-level RBAC plus attribute-based conditions on time, location, sensitivity, and anomaly signals.

  4. 04

    Sandboxed the small language model with no database access, no tools, and no memory across requests, then validated every output back against its input claims to catch hallucination.

  5. 05

    Logged every query, policy decision, and access event to an immutable audit trail, and mapped the whole system to BFSI, healthcare, and government compliance.

The guarantee

A deterministic, sub-second pipeline that answers in natural language while the model layer stays fully untrusted. The sensitive part of the system never sees the sensitive data.

Enforced by

  • Sandboxed SLM: no data, no tools, no memory

  • Claim-based immutable data model with provenance

  • RBAC + ABAC with department-level silos

  • Output validation against source claims

  • Immutable, queryable audit trail

  • Maps to RBI FREE-AI, DPDP, SEBI, GDPR, SOC 2, ISO 27001, HIPAA