Sakshi is not another model or another dashboard, it's the accountable record beneath the agents you already run. Every agent inventoried, every decision witnessed, every action bounded, every regulator mapped, inside your own environment.
The Sakshi console, live, agent inventory, decision volume, AI-disclosure coverage, and a chain that recomputes and verifies itself.
Predict & prevent
The five modules record and bound. On top of them, Sakshi runs the lifecycle regulators actually ask for: certify an agent before it ships, watch it in production, and catch the day it silently changes — continuous monitoring, kill switch, red-teaming and independent validation, RBI-mapped.
No agent ships without a governance sign-off. Run the candidate through the certification battery — adversarial red-team, fairness, evaluation gates, a formally-verified autonomy policy, verified identity — for a signed, RBI-mapped certificate with an expiry and re-certification triggers.
A live monitor over the whole fleet that halts, escalates or steers the moment an agent drifts, is poisoned, or goes off-policy — using the same drilled kill switch a human would, every intervention on the tamper-evident record.
Agents change silently — a model swap, a prompt injection, drift. Sakshi fingerprints each agent's behaviour and tells you the day it started behaving differently, and whether the change was declared or silent — which automatically sends it back through certification.
How it fits
Your agents keep running where they run. Sakshi observes, governs, and records, and never sees a raw identifier.
An agent that isn't in the inventory can't record a decision. Register is the live roster of every agent and model in production: who owns it, what it's allowed to do, and where it came from.
Every decision an agent makes is captured and hashed into a per-tenant chain: context, model identity, the reasoning steps, the action, and who touched it. PII is tokenized before anything is stored, and any record can be recomputed to prove it wasn't altered.
Envelopes decide, per action, whether an agent may act on its own, must route to a human, or must stop, by stakes, confidence, and blast radius. Oversight telemetry catches rubber-stamping, and a drilled kill switch halts a fleet the moment it drifts.
Every instrument, RBI MRM, DPDP, SEBI Reg 16C, IRDAI, is encoded as data and scored against your live evidence. A clause is "ready" only when a real check backs it, so the readiness score is honest, leverage-ordered, and directly actionable.
Bias testing that speaks Indian categories, gender (including transgender) and caste , computed off the live decision stream. Declared attributes wherever they exist; flags only fire on a real threshold breach and statistical significance, so the screen doesn't cry wolf.
The spine
Every record links to the one before it by hash. Anyone, your auditor, your regulator, can recompute the whole chain in your environment and confirm nothing was altered. Change a single field and the chain names exactly where.
That's the difference between a log you're asked to believe and evidence that proves itself.
A few lines with the Python SDK, or zero code through the MCP server and interception proxy. Framework middleware for OpenAI, Anthropic, Gemini, Bedrock, LangGraph and MCP.
Developer docsSingle-tenant, installs in your VPC or on-prem, or Rota-managed in an India region. Evidence never leaves your environment; nightly backups; one-command install.
Hosting & residencyExplore the console with synthetic data, or get your own private sandbox in minutes.
