[bmdpat]

§ 001 / THE LAB COURSE

The Local AI Engineering Lab

invite-only beta

Eight build-first modules. Every open model runs on hardware you control, and every module ends in a benchmark you publish.

8 modules~1.8 hrs reading~14.5 hrs hands-onBuild-first, benchmark-gated

Connecting sign-in... If this never resolves, sign-in is misconfigured or blocked. Email pat@bmdpat.com and mention the course.

§ 002 / HOW THIS TEACHES

After the eight builds you can

  • Train a language model from a blank file and read its loss curve like an instrument
  • Fine-tune an open model on your own data with LoRA and prove the lift with an eval
  • Build retrieval, agent, and reasoning pipelines that run on hardware you control
  • Harden an agent with memory, injection defense, and runtime budget guards
  • Publish benchmark-backed build logs that double as a portfolio

Build first

Every section starts with something to make, and theory arrives when the build needs it. Struggling with a problem before the explanation measurably improves transfer.

Check yourself

Sections end with retrieval questions and every module ends with a closed-book rebuild. Self-testing is the strongest replicated result in the learning literature.

Spaced by design

Each module reuses the previous module's artifact, so old material gets recalled weeks later exactly when forgetting would set in.

Prove it in public

Every module ends in a published benchmark. Explaining your build to strangers is the Feynman step, and the artifact is your proof-of-build.

§ 003 / CURRICULUM

  1. FOUNDATIONS24m read · 120m build

    LLM Playground on Local Hardware

    Build a language model from nothing and learn why every piece exists

    Proof: A decoding-strategy benchmark table, quality vs tokens/sec on your card.

  2. ADAPTATION12m read · 105m build

    Fine-Tune Llama 3.1 8B into a Support Bot

    LoRA on one consumer GPU, measured before and after

    Proof: A before/after eval table plus the VRAM and wall-clock cost of the tune.

  3. RETRIEVAL11m read · 100m build

    Build a Local RAG Pipeline

    Chunk, embed, retrieve, and guard, all on your box

    Proof: A retrieval-quality table comparing three chunking strategies.

  4. AGENTS11m read · 110m build

    Build an Ask-the-Web Agent

    Routing, reflection, and tools over MCP, priced against the API

    Proof: A local-vs-API cost curve for the same agent trace.

  5. REASONING11m read · 105m build

    Add a Deep Research Capability

    Spend inference-time compute on a local reasoning model, and measure it

    Proof: An inference-time-scaling curve, answer quality vs thinking tokens.

  6. GENERATION12m read · 95m build

    Ship a Local Diffusion Image Service

    Text-to-image behind a small API, benchmarked across quant levels

    Proof: A VRAM/latency/quality grid across quant levels for the same prompt.

  7. OBSERVABILITY13m read · 115m build

    Harden the Agent - Memory, Guards, and Observability

    Make Module 4's agent safe to leave running

    Proof: A red-team log, injection attacks attempted vs which guard caught them.

  8. CAPSTONE11m read · 120m build

    Capstone - Ship One Production-Shaped Local Agent

    Compose the seven parts into one system a stranger can run

    Proof: A public repo plus a lab report with at least one real benchmark table.

§ 004 / COMPUTE

You do not need to buy anything to take this course. 4 ways to run the work, no card required:

Your browser

free

Every module ships a rebuild exercise that runs right on the page (Python compiled to WebAssembly). Nothing to install, nothing uploaded.

Your own GPU

hardware you already own

The lab's home turf. One consumer card (8GB+ VRAM) runs every module; the writeups are tuned for local rigs first.

NVIDIA free endpoints

free credits

No GPU? build.nvidia.com serves Llama and DeepSeek-R1 behind a free OpenAI-compatible API. The agent and reasoning modules (4, 5, 7) run against it with a one-line base-url change.

Colab free tier

free T4 GPU

The training modules (1 and 2) fit Colab's free T4 when you don't have a local card. Same code, slower card, still yours to run.

The Local AI Engineering Lab | Patrick Hughes