library / AI Lab / AI Learning Tracks / llm-builder-zero-to-shipped-agentreviewed 2026-08-12

AI Lab / AI Learning Tracks

LLM Builder: Zero to Shipped Agent

A practical path from prompting to RAG, evals, tools, agents, deployment, and failure analysis.

Room context

Learning track

Build ladder.

Ship one useful AI assistant with a clear task, test set, retrieval boundary, and failure log.

Prerequisites

  • basic Python
  • basic web literacy
  • willingness to test with boring examples
  1. 01
    Prompting as interface design

    Write prompts with role, task, context, output contract, and examples.

    Turn a messy note into a structured student resource.
  2. 02
    RAG without superstition

    Understand chunks, retrieval, citations, stale data, and source trust.

    Build a scholarship finder over a small verified corpus.
  3. 03
    Evals before vibes

    Create small pass/fail cases and regression examples.

    Evaluate the scholarship finder on 20 realistic questions.
  4. 04
    Tools and agents

    Use tools only where the model needs external action or exact computation.

    Add a deadline checker and source verifier.
  5. 05
    Ship and observe

    Deploy with logs, known limitations, and a feedback loop.

    Publish a read-only demo with a failure diary.

Checkpoints

  • Can explain what the model knows
  • Can show citations
  • Can run evals
  • Can name failure modes
  • Can improve from feedback

Public Promise

This track is not a tour of AI buzzwords. The promised artifact is one small assistant that answers a real question, cites what it used, logs what failed, and survives a second look.

The useful habit is to ask for evidence before adding agentic sparkle. A boring eval file is usually more honest than a dramatic demo.

  • task
  • source boundary
  • test set
  • retrieval rule
  • failure log
  • shipping boundary

What To Build

A good first build is narrow: a scholarship helper over verified sources, a student FAQ assistant, or a small research-note organizer. Keep the first version read-only until source, privacy, and failure modes are visible.

Each module should end in a commit, a screenshot, a failing example, and one paragraph explaining what changed in your judgment.

Nearby

You might go next