Work

AI platformsProduction2025 - Present

A consumer-facing AI assistant on Amazon Bedrock AgentCore

Agentic assistant with grounded retrieval, governed tool access, and conversation memory, plus the shared AI platform internal tools now build on.

  1. Client

    • Consumer app

      Chat UI

  2. Agent

    • AgentCore Runtime

      LangGraph agent in Python

  3. Context

    • PostgreSQL + pgvector

      Retrieval over company knowledge

    • Redis

      Conversation state and cache

  4. Actions

    • AgentCore Gateway

      Governed tool access

    • Internal services

      Exposed as agent tools

Request path through the assistant platform.

Context

I build GenAI and agentic solutions across the company on Amazon Bedrock AgentCore and in-house AI infrastructure: chatbots, agentic workflows, and AI-powered tools for diverse use cases. The flagship is a conversational assistant that consumers use directly, which raises the bar on answer quality, latency, safety, and cost compared with an internal tool.

The problem

A consumer assistant has to answer from trusted company knowledge, call real systems on the user's behalf without overreaching, keep context across a conversation, and stay predictable under load. A single prompt over a model API does none of that reliably.

My role

Hands-on engineer and architect: I build the agents and workflows, and shape the architecture and design of the underlying AI tooling and infrastructure stacks so the team can ship scalable, secure, production-ready AI platforms.

Approach

  • Agents run on Amazon Bedrock AgentCore Runtime, built in Python with LangChain and LangGraph, so orchestration, tool calling, and model access sit in a managed, isolated runtime instead of hand-rolled serving code.
  • Tools follow the Model Context Protocol (MCP), so the same capabilities can be reused across assistants and agentic workflows.
  • Tools are exposed through AgentCore Gateway, which gives the agent one governed entry point to internal APIs instead of per-agent integrations.
  • Retrieval uses PostgreSQL with pgvector as the vector store, so embeddings sit next to the relational data they describe and share its operations and access model.
  • Redis holds conversation state and caches hot lookups, keeping multi-turn context and repeat queries fast.
  • The same building blocks are being reused for internal AI tools, so each new assistant starts from a proven retrieval, tool, and memory layer.

Outcome

  • One platform pattern serving both a consumer product and a growing set of internal AI tools.
  • Tool access governed centrally through the gateway rather than per agent.
  • Vector search kept inside PostgreSQL, avoiding a separate vector database to operate and secure.
Next storyCloud-native microservices for mortgage loan automation