AgentGate
|Founding Engineer - AI Agent Guardrails Platform
Bengaluru, Karnataka, India
Summary
Led the development of a framework-agnostic AI agent guardrails platform, establishing critical enforcement layers and pipelines to ensure secure, policy-compliant operations across diverse digital surfaces.
Highlights
Architected and implemented a dual-surface, framework-agnostic enforcement layer returning allow/block/require-approval, integrating with SecureBrowser (Playwright) and MCP gateway for agent tool calls to downstream services (payments, code hosting, databases).
Developed a high-performance ingestion and evaluation pipeline that normalized intercepted actions into a uniform ActionEvent, running a cheapest-first cascade across 200+ terms and 13 categories.
Optimized the evaluation pipeline using phonetic/stemming and on-device bge-small embeddings (~67 MB, CPU, ~2 ms latency, no LLM in loop), effectively grading actions from CRITICAL to LOW against YAML policies.
Designed and built a scalable per-session multi-agent coordination system on a single host, incorporating role-based permissions, per-agent/shared rate limits, and an append-only hash-chained evidence system for horizontal scale-out via Redis-backed distributed counters and locks.