The AI drafts the exam question. A specialist decides.

A prototype that grew into a multi-agent LangGraph platform at ACT, now in production and rolling out across subject areas to speed up content creation.

Role
Senior full-stack engineer
When
Sep 2025 – Aug 2026
Core stack
LangGraph, NestJS, Next.js, AWS
Status
In production

The story

  1. Problem

    Every new item cost a specialist hours

    ACT writes exams that 2M+ students take each year. Each question was drafted, checked against the question bank, and reviewed by hand.

  2. Prototype

    Prove one workflow first

    We mapped the work with specialists and stakeholders, then built one thing: a prompt chain that drafts a first version of an item. It was approved.

  3. Production

    Grow it into a platform

    Subject agents, passage generation, five parallel reviewers, and a duplicate check, deployed to AWS. I owned features from the agents to the UI.

How one question gets made

Solid blocks are where a person decides. Dashed lines send work back.

Specialista personAgentsLangGraphPlatformNestJS, AWSRequesta new itemSchedulelogged jobGenerateby subjectBank checkRAG searchReview ×5in parallelSynthesizefix listSign-offhuman gateQTI exportitem banks80%+ similarchanges leftSpecialista personAgentsLangGraphPlatformNestJS, AWSRequesta new itemSchedulelogged jobGenerateby subjectBank checkRAG searchReview ×5in parallelSynthesizefix listSign-offhuman gateQTI exportitem banks80%+ similarchanges left

Calls I made

  • Small agents, not one mega-prompt

    Question writing, passage writing, and review each have their own prompt and rules. A new subject is a new combination, not a rewrite.

  • A cheap retry before a human

    An item 80% similar or more to one in the bank is regenerated once. Only a second miss takes a specialist’s time.

  • Resume from checkpoints, never from zero

    Long jobs resume from their last LangGraph checkpoint after a crash or a deploy, so tokens already spent aren’t spent twice.

Built end to end

Product
  • Next.js app for specialists
  • Live progress over authenticated Socket.IO
API
  • NestJS
  • Collision-free job scheduling
  • Audit log of every job
Agents
  • LangGraph flows
  • Durable checkpoints
  • Subject specialists
Data
  • RAG over the question bank
  • RAG over textbook PDFs
  • IMS QTI export
Delivery
  • Docker
  • Kubernetes on AWS
  • Jenkins CI/CD
  • Jest and Playwright concurrency tests

“Luis was easy to work with: skilled, collaborative, and responsive. He listened well, found good solutions, and met commitments. I’d work with him again in a heartbeat.”

Lance Blackstone
Product Owner @ ACT
Next case · SoloDevShip features with an AI team, not a single agent