Skip to content
Hire us

Learn to ship AI.Not just talk about it.

A hands-on path from model fundamentals to RAG, agents, evaluation, and production architecture—built around the decisions real AI engineers make.

View full curriculum

No sign-up. Pick up where you left off.

8 modules · 3 real builds · progress saved

One useful next step.

Your work is stored in this browser, so the dashboard stays honest—no invented streaks or fake completion.

View full curriculum

Continue here55 minBeginner

01 · Foundation

AI systems, not AI magic

Learn how modern AI products turn inputs, models, context, tools, and evaluation into a dependable system.

  • Map an AI request from input to output
  • Separate deterministic code from model judgment
  • Define a measurable success metric

55 min · includes a build lab

Your progress

0%

Course progress
0%
Modules
0/8
Learned
0h
Remaining
8

Your next three sessions

Build proof, not another certificate.

Three scoped portfolio projects include milestones, acceptance criteria, and the evidence an engineering team wants to see.

Explore build lab

Curriculum

Module 01FoundationBeginner55 min

AI systems, not AI magic

Learn how modern AI products turn inputs, models, context, tools, and evaluation into a dependable system.

The mental models that matter

  1. 01
    The system view

    An AI feature is a pipeline. The model is only one part alongside data, instructions, tools, validation, and feedback.

  2. 02
    Where models help

    Use models for fuzzy language tasks: extraction, classification, generation, ranking, and semantic search.

  3. 03
    Where code wins

    Use ordinary code for permissions, calculations, state changes, billing, and rules that must behave predictably.

Design a support triage system

Turn an incoming support message into a category, urgency score, suggested reply, and human-review decision.

Definition of done

  • Input/output contract
  • Four edge cases
  • One quality metric
  • Human escalation rule
starter.js
const ticket = await classify({
  message,
  output: { category: "enum", urgency: "1-5" }
});

if (ticket.urgency >= 4) await requestHumanReview(ticket);

Knowledge check

Which part should never rely on a model guess?

After this module

  • Map an AI request from input to output
  • Separate deterministic code from model judgment
  • Define a measurable success metric

Saved automatically in this browser.

Projects with a real definition of done.

Each brief is intentionally scoped to produce an inspectable engineering artifact—not another shallow chatbot clone.

  1. Project 01IntermediateRAG · 2 weeks

    Source-grounded knowledge assistant

    A document assistant that refuses unsupported answers and cites the exact passage it used.

    • Next.js
    • Postgres
    • Vector search
    • LLM API

    Build milestones

    1. Ingest and chunk three document formats
    2. Retrieve and rerank relevant evidence
    3. Generate cited answers with a no-answer path
    4. Evaluate retrieval and faithfulness on 30 questions

    Proof of work

    Demo + architecture README + evaluation report

  2. Project 02AdvancedAgents · 3 weeks

    Tool-using support operator

    An agent that researches an account, drafts a resolution, and asks before making any customer-facing change.

    • Tool schemas
    • MCP
    • Approval UI
    • Trace logs

    Build milestones

    1. Create read-only account and policy tools
    2. Add planning and bounded retry behavior
    3. Gate refunds and outbound messages
    4. Replay 20 traces and classify failures

    Proof of work

    Working trace viewer + safety matrix + demo

  3. Project 03AdvancedEvals · 2 weeks

    AI release evaluation console

    A dashboard that compares prompt or model versions and blocks regressions before deployment.

    • Golden dataset
    • Model graders
    • CI checks
    • Analytics

    Build milestones

    1. Collect representative test cases
    2. Add exact, rubric, and safety graders
    3. Compare versions by slice
    4. Create a pass/fail release gate

    Proof of work

    Version comparison + CI report + failure taxonomy

Complete 3 more modules before starting the first build.

Go deeper with primary sources.

A short shelf of documentation and code-first courses. Better than a giant list you will never open.

Plain-English lookup

Context window
The input and conversation history a model can consider for one response.
Embedding
A numeric representation that places semantically similar content near each other.
Evaluation
A repeatable test that measures model or system behavior on representative inputs.
Function calling
A structured way for a model to request an application-defined operation.
Grounding
Connecting an answer to trusted evidence supplied at response time.
Hallucination
Content that sounds plausible but is unsupported or incorrect.
MCP
A protocol for exposing tools, resources, and prompts to compatible AI clients.
RAG
Retrieving relevant evidence before asking a model to generate an answer.