pedro castanheiraPT

Pedro Castanheira Costa · AI Engineer · Tech Lead

I build AI agents that run in production — from the data to deployment and evaluation.

I am the technical lead of a squad responsible for five AI products in production at an education company. I work from problem discovery with business teams through architecture, implementation, observability and evaluation — mostly with Python, LangGraph and FastAPI, and TypeScript with NestJS on the backend.

spreadsheets, PDFs,internal systemsmessy datacurationreliablecontextagents(LangGraph)tools · MCPmemoryproduction24/7evals in CI+ Langfusemeasure and adjust↑ this is what I build, end to end
  • 70–120sales a day closed by a conversational agent, 24/7
  • ~140ksubmissions graded per month by an async pipeline with LangChain
  • ~45%lower OCR cost by discarding invalid documents before extraction
See the workGitHubLinkedInGet in touch

Work

In production

Three of the five products I lead technically. The names stay out; the problem, the decisions and the numbers stay in.

  • Production

    Sales and enrollment agent

    Built a conversational LangGraph agent with a router, three subagents and tools integrated with internal systems — it answers course questions and drives enrollment, 24/7.

    70–120sales a day

  • Production

    Assignment grading at scale

    Designed an asynchronous NestJS pipeline on BullMQ that syncs submissions from the LMS, sends each one to LangChain grading and writes the grade and feedback back to Moodle.

    ~140ksubmissions graded per month

  • Production

    Document triage before OCR

    Built a two-stage visual validator with ConvNeXt-Tiny that discards illegible or wrong-type documents before extraction with Document AI.

    ~45%lower OCR cost

Personal projects

Where I test ideas in the open: graph-based code review, RAG evaluation and tooling for coding agents.

  • Personal · open source

    Cast Review

    Built a pull-request review platform that indexes the repository as a Neo4j graph and exposes indexing, context and analysis through an MCP server.

    5tools on the MCP server

  • Personal · open source

    Groundtruth — an evaluation gate for RAG

    Built a harness that measures RAG changes before merge and blocks the PR when Recall@10 drops by more than 1 percentage point.

    1 ppRecall@10 drop is enough to block a merge

  • Personal · open source

    Cast Code

    Built a multi-agent coding CLI that plans, delegates to specialists, uses tools over MCP and checks its own work in the terminal.

    ~30MCP integration templates

  • Personal · open source

    Cast Skills

    Created an npm package that installs project-context skills and a spec-driven workflow into six AI coding tools.

    6AI tools supported

Evidence

Every skill here points to proof.

Instead of a list of technologies, a graph: each skill links to the projects where it was used. Hover, tap or use the keyboard to light up the connections.

Pick a skill ↑

Method

How an AI product gets built with me.

The loop I run with my squad. Evaluation and observability feed back into discovery, so the product improves on evidence — not on impressions.

  1. discovery

    Diagnosis with business teams: which decision or task changes, and how we will know it worked.

  2. curation

    Understand documents and spreadsheets, relate entities and metadata, decide what the source of truth is.

  3. context

    Chunking, hybrid search, context routing and short- and long-term memory.

  4. agents

    LangGraph graphs, subagents, tools, API integrations and MCP servers.

  5. evaluation

    Evaluators plus retrieval and full-graph tests running in CI/CD.

  6. operation

    Langfuse, cost, error and hallucination metrics in production.

Path

Where this experience comes from.

  1. Aug 2025 — present

    Software Engineer — AI & GenAI · Squad Tech Lead

    Inova Soluções Educacionais

    • Technical lead of the squad and of five AI products in production: diagnosis with business teams, architecture, deployment and metrics.
    • Built LangGraph agents, a grading pipeline on BullMQ and LangChain, a multi-tenant RAG on Qdrant for tutoring and ConvNeXt document classifiers.
    • Rolled out LLMOps with Langfuse and evaluators; CI/CD with retrieval and full-graph evaluation, plus cost, error and hallucination metrics.

  2. May 2025 — Aug 2025

    Data intern

    Centro Universitário UniFatecie

    • Power BI dashboards and queries for Jasper Reports.

  3. Mar 2025 — May 2025

    Front-end intern

    Vertti Tecnologia

    • Built features in React and JavaScript.

Contact

Let's talk about agents, data and AI products.

For conversations about AI engineering, agents or RAG, email or LinkedIn are the best ways to reach me.

LinkedInGitHub

Maringá, Brazil