CH
WorkAboutArticlesLab
kikeztw@gmail.comen/es

Let's build something

CH

© 2026 Carlos Hurtado · All rights reserved.

← Back

Ximple Tech

Agentic AI Platform

End-to-end architecture for LLM agents with tool-use, MCP servers and RAG over pgvector & Pinecone, deployed on AWS serverless.

  • #AGENTS
  • #RAG
  • #MCP
  • #AWS
Role
AI Software Engineer, Ximple Tech
Period
2025 – 2026
pgvectorPineconeMCP · WhatsAppGoogle APIsSQS · Step Fnagenttool-use · LLM

The challenge

Design and own the end-to-end architecture of an AI platform: LLM agents that call tools, connect to internal and external systems, and answer from private knowledge.

Approach

  1. 01Defined services, domain boundaries and contracts, and deployed them on AWS (serverless plus EC2).
  2. 02Built LLM agents with tool-use on OpenAI and the Vercel AI SDK, with Zod-validated tool schemas and robust error handling.
  3. 03Built MCP servers and orchestration components to plug internal and external tools into the agents.
  4. 04Implemented RAG and semantic search with two strategies, PostgreSQL + pgvector and Pinecone: ingestion, chunking, embeddings and contextual retrieval.
  5. 05Wired agent channels and automations through Pipedream and WhatsApp-style messaging, connected to internal services via MCP.

Stack

  • TypeScript
  • Node.js
  • OpenAI
  • Vercel AI SDK
  • Zod
  • MCP
  • PostgreSQL
  • pgvector
  • Pinecone
  • Prisma
  • AWS Lambda
  • SQS
  • Step Functions
  • NestJS
  • GraphQL

Outcomes

  • One platform covering agents, tool integrations and retrieval, owned from architecture to deployment.
  • Semantic search implemented with two retrieval strategies: PostgreSQL + pgvector and Pinecone.
  • Document ingestion pipeline that turns PDFs into searchable knowledge.
← Previous caseCommunity AutomationNext case →Mobile Engineering