production AI agents · agent security · multi-agent systems

I build secure AI agents and AI-powered security products.

Senior software engineer at AWS with 10+ years building complex products from concept to production. I built a production AI agent — a first prototype in 6 weeks, now in pilot — and I build agents secure from the first prototype: prompt-injection defense, agent threat modeling, and Bedrock Guardrails baked in. Ask me anything; answers are grounded in my resume.

Popular questions
View work →
6 weeks
Text2SQL agent on Strands — 40% ahead of a 10-week estimate
12,000+ users
Cloud platform migration delivered
10-agent
SDLC pipeline distributed across Amazon teams
Focus

Where I work.

01
Production AI agents
Taking agents from prototype to production on Amazon Bedrock — MCP Server and AgentCore Runtime (Strands), Text2SQL, vector search, with evals and observability built in.
02
Agent security
Prompt-injection defense, memory-poisoning safeguards, and agent threat modeling — plus application security reviews and penetration testing. Security from the first prototype, not bolted on.
03
Multi-agent systems
Composable skills, agent SOPs, and handoff protocols — like a 10-agent SDLC pipeline distributed company-wide. Plus proposed protocols for agent governance and state portability (AGP · DAP).
Selected work

Projects.

Production agents, the security around them, and the systems they run on. Click any project to expand. Research items are marked proposed.

In pilot with the operations team (~50 users), with rollout to 12,000+ AWS sellers underway over the next ~6 months. Led the technical design as lead developer. Integrated LLMs through Amazon Bedrock with dual paths — MCP Server and AgentCore Runtime (Strands) — plus Text2SQL and vector search. Security owned end to end: Bedrock Guardrails, prompt-injection filters, and memory-poisoning safeguards, with formal application-security approval before launch.

~50 users
pilot with the operations team
12,000+
sellers — full rollout over ~6 months
6 weeks
to prototype (40% faster than a 10-week estimate)
sub-15s
response times
Amazon BedrockMCP ServerAgentCore · StrandsText2SQLvector search
Interactive

Explore, hands-on.

The same interactive views the chat will render inline as MCP Apps in Phase 2 — live here first. AGP & DAP are proposed / conceptual, not implemented.

App view · ui://resume.explorer
⛨ sandboxed · ready
Amazon Web Services · 2016 — present
  • Built BOB, a production AI agent for AWS sellers (compensation and policy questions in natural language) — in pilot with the operations team (~50 users), expanding to 12,000+ sellers over ~6 months.
  • Delivered the first prototype in 6 weeks — 40% ahead of a 10-week estimate — then iterated toward rollout.
  • Integrated LLMs through Amazon Bedrock with dual paths — MCP Server and AgentCore Runtime (Strands) — plus Text2SQL and vector search.
  • Owned security, evaluation, and observability: Bedrock Guardrails, prompt-injection filters, and memory-poisoning safeguards, with formal application-security approval before launch.
  • Built a 10-agent SDLC pipeline on Kiro using AgentCore Runtime and Strands, distributed company-wide for reuse across Amazon teams.
  • Led application security reviews and penetration testing for the Varicent SaaS migration for 12,000+ users.
Amazon BedrockMCP ServerAgentCore · StrandsText2SQLvector search
About

Secure AI agents
you can audit.

I'm a senior software engineer at AWS with 10+ years shipping complex products from concept to production. Most recently I built a production AI agent on Amazon Bedrock with MCP and AgentCore — a first prototype in 6 weeks (40% ahead of estimate), now in pilot — with security owned end to end.

I work where applied AI, product, and security meet: rapid 0-to-1 prototyping, agent threat modeling and application security reviews, and the data and systems plumbing — ETL, Text2SQL, evals — that production agents actually need. I like ambiguous problems, fast iteration, and staying close to the people who use what I build.

Srinivasan Nambi
secure AI agents
Principles
Secure by design
Security from the first prototype — prompt-injection defense, threat modeling, application security reviews.
Zero to one, fast
Prototype to production in weeks. At home in ambiguity, rapid iteration, and shipping end to end.
Auditable & honest
Every decision traceable to its source — and clear about what shipped versus what's proposed.

Let's talk.

Whether you're a researcher, founder, or hiring — if you're working on hard problems in secure, production AI agents, I'd love to hear from you.
View CV ↗
LinkedIn →