I build systems for high-trust AI: deterministic orchestration, policy enforcement, and cryptographically-verifiable forensics.

My work focuses on replacing unreliable "Generative Judges" with deterministic NLI Cross-Encoders and basic hash chains with Ed25519 asymmetric signatures.

Deterministic RAG Verification & Governance A reference implementation for local-first AI that prioritizes grounding integrity over marketing hype.

• Innovation: Replaced slow LLM judges (~2000ms) with a DeBERTa-v3 NLI gate (80ms).
• Security: Full Ed25519 forensic chain for non-repudiable audit trails.
• Trust: Hardware-backed key management (OS Keyring/TPM integration).

The core retrieval engine for sovereign environments. Hybrid search (BM25 + Vector) with strict ABAC policy gating and signed forensics.

A lightweight, auditable orchestration layer for local AI agents, focusing on fail-closed tool governance and tamper-evident execution logs.

• Grounding Verification: Using Natural Language Inference (NLI) for deterministic hallucination control.
• Asymmetric Forensics: Implementing Ed25519 signatures in audit logs to ensure non-repudiation.
• Local-First Governance: Policy enforcement (ABAC) at the retrieval and tool-call layers without external dependencies.
• Hardware-Backed Trust: Integrating with OS-level secure enclaves for identity and signature management.

I am an advocate for open standards in AI identity and governance.

• IETF Draft: Attested Agent Identity (PTV)
• Vision: Transitioning AI governance from "Trust but Verify" to "Verify then Trust."

• LinkedIn: linkedin.com/in/anandkrshnn
• Email: ananda.krishnan@hotmail.com
• PyPI: sovereign-ai-stack

Building for healthcare, finance, and high-stakes infrastructure where "good enough" AI isn't an option.