|
I'm Arjun — an AI/ML engineer and governance architect in Bharat 🇮🇳, building the trust infrastructure responsible AI needs. Deploying AI for fintech, HR and healthtech made one thing clear: enforcement was missing. Systems were intelligent; none were accountable. That became Kavach: authorization for AI agents and automated decisions, where authority comes from a business event, not from a model. What drives me:
|
struct TheIndicSentinel {
name: "Arjun",
mission: "Digital armor for AI",
now: "Kavach (Rust core)",
principle: "Not just intelligent. Accountable.",
} |
Authorization and runtime control for AI agents. Free and open source (Apache-2.0). Pre-release: not production-ready, no external security review yet.
An agent should not hold the credential to act unless it is authorised to act. Every action needs a signed Task Mandate issued from a system-of-record event, is checked against Cedar policy, trusted time and server-side counters, is recorded as signed evidence before it runs, and only then gets a short-lived credential bound to that exact request.
| Area | Status |
|---|---|
| Task Mandates (issue, delegate, revoke) | ✅ |
| Cedar agent policies, formally analysed in CI (cvc5) | ✅ |
| Signed tool registry, reference-only parameters | ✅ |
| Signed, hash-chained evidence, committed before execution | ✅ |
| Short-lived credentials (JWS in JWE, ≤ 15 s) | ✅ |
| Gateway, acceptance suite, network isolation (deployed stack) | ✅ built, still pre-release |
| Signed evidence checkpoints, export, offline verifier | 🚧 in progress |
Developer CLI: kavach init, demo, authorize, why, attack |
🗺️ v0.1 preview |
Agent ──► Gateway ──► Mandate ► Cedar policy ► Trusted time ► Counters
│
▼
Signed evidence (before action)
│
▼
Short-lived credential ──► Provider ──► Signed outcome
| Phase | Focus |
|---|---|
| Now | Evidence checkpoints, export and offline verification |
| v0.1 preview | Developer CLI first: kavach init, demo, offline authorize, why, attack; supply-chain hardening, benchmarks, fuzzing, open KMS/HSM adapter |
| v0.2 | kavach studio terminal UI (read-only, no telemetry) |
| Later | Approvals and step-up, taint tracking, MCP adapter, RBI/DPDP evidence packs, multi-tenant and HA reference, external security review |
An open, auditable control layer between AI agents and the systems they touch, so regulated Indian lenders can prove what an agent was allowed to do and why. Security properties stay free; an optional enterprise plane (fleet management, SSO, audit-export packs) lives in a separate repo. See OPEN_CORE.md.
| Project | What it is |
|---|---|
| KavachX v2 | Earlier Python governance engine: risk scoring, safety classifiers, audit chain |
| dpdpa-pii-scrubber | Zero-dependency redaction of Aadhaar, PAN, UPI, IFSC and other Indian identifiers |
| Awesome AI Governance Bharat | Curated laws, frameworks, tools and research for Indian AI governance |
| VyaparGPT | Hindi and English assistant for Indian SMEs |
Building in public. Reach out if you work on AI safety, Indian AI policy or responsible ML infrastructure.
🛡️ Kavach — Accountable AI infrastructure · Built with conviction in 🇮🇳 India