Why Now
Your Regulator Is Ahead of Your Tooling
The RBI's draft Model Risk Management directions and the DPDP Act ask for proof, not policy documents. Sakshi produces that proof as a by-product of running your agents.
Evidence, Not Assertions
Attested drills, signed evidence packs, decision chains that verify. If it isn't recorded, Sakshi doesn't claim it.
India-First by Design
DPDP-aware PII handling, Hinglish guardrails, fairness screens for Indian demographics — core architecture, not localization.
Sovereign Deployment
Single-tenant, in your environment. With local models, a witnessed decision sends zero bytes outside your infrastructure.
The Platform
Five Modules, One Chain of Evidence
Register
Know Your Agents — every agent and model inventoried, each with an accountable human owner.
Witness
A tamper-evident flight recorder for agent decisions, with PII tokenized at ingest.
Bound
Bounded autonomy — policy envelopes, human review queues, and a drilled kill switch.
Vidhi
Signed evidence packs mapped clause-by-clause to the RBI draft MRM directions and the DPDP Act.
Vishwas
Fairness you can cite — declared-attribute bias screens and pre-deployment fairness runs.
Integration
A Few Lines of Code
Instrument an existing agent in minutes — governance should never be a migration project.
from sakshi import SakshiClient
sakshi = SakshiClient("https://sakshi.yourbank.in",
api_key=key)
agent_id = sakshi.register("loan-decision-agent",
owner_name="R. Sharma",
owner_email="[email protected]")
with sakshi.witness(agent_id) as record:
record.step("assess", score=0.81)
decision = sakshi.enforce(agent_id,
"approve_loan", stakes=0.7)
record.action(outcome="approved")SDK — available now
Open source on PyPI. Capture fails open; enforcement fails closed.
Gateway — in design
Zero code: point your OpenAI-compatible base URL at Sakshi and every call is governed at the wire.
In the Open
Built in Public, Where It Can Be
RBI's final MRM directions are coming.
Be ready with evidence, not assertions.