$ ls curriculum/ --depth=deep

Ten core modules you couldn't have just googled

Built from primary sources and taught at mechanism depth — covering the misconceptions most tutorials skip. Self-paced, and unlocked when you enroll. Lesson content is still being written — the lifecycle below is complete.

$ cat ./lifecycle

01 → 02

Source & run

Read model cards on HF, run GGUF locally, expose a local API.

03 → 05

Prompt, ground & tools

Chat templates and sampling, RAG retrieval, then tool-use agent loops.

06 → 08

Data, tune & evaluate

Curate datasets, LoRA/QLoRA → GGUF with Unsloth, then score on an eval harness.

09 → 10

Optimize & ship

imatrix Q4_K_M quants, serve for throughput, package and hand off.

$ ls -la curriculum/

M01rwx

Local-AI Foundations

01-foundations.md

Why local, the bandwidth truth, and the toolchain mpl installs.

M02rwx

Running Open Models

02-run.md

Pull, run, and serve open models locally.

M03rwx

Prompting & Sampling

03-prompt.md

Chat templates, the sampler pipeline, and constrained decoding.

M04rwx

Grounding & RAG

04-ground.md

Embeddings, vector search, and the retrieval-augmented generation pipeline.

M05rwx

Tools & Agents

05-tools.md

Tool-calling over the serving API, and building a local agent loop from primitives.

M06rwx

Data & Datasets

06-data.md

The chat/completion shapes mpl finetune consumes, train/val splits, and contamination.

M07rwx

Fine-Tuning (LoRA/QLoRA)

07-finetune.md

LoRA rank/alpha math, the QLoRA 4-bit memory budget, and the fuse → GGUF pipeline.

M08rwx

Evaluation

08-evaluate.md

The mpl eval scorer, the MCQ extraction failure mode, and regression sets.

M09rwx

Optimization & Serving

09-optimize.md

PTQ quant types and the imatrix, plus serving for throughput with batching.

M10rwx

Shipping a Local-AI App

10-ship.md

Packaging by composition and clean teardown via the install ledger.

$ ls specializations/

Then specialize: three tracks past the core

The core feeds the path from using AI to building it. Automate real work with Agent Harness Engineering — the runtime layer around a local model — unlocked at module 05. Then Build the model itself with Inference Engineering or a model from scratch, both stacked on the full core. Production Local AI is planned.

track D · 12 modulesstart here

Agent Harness Engineering

Build the runtime around a local model: context, memory, tools, permissions, evals, and traces. Start by turning a model into a streaming chat loop.

unlocks after module 05

track A · 8 modulesmajor

Inference Engineering

Make local inference fast: profiling, attention/KV optimization, quantization, speculative decoding, and production engines.

unlocks after the ten-module core

2 cloud labs — run on a box you launch

track B · 12 modulesmajor

Build a Model from Scratch

Build a small foundation model: the decoder, latent attention, mixture-of-experts, FP8, training, and post-training.

unlocks after the ten-module core

1 cloud lab — runs on a box you launch

track C · plannedcoming

Production Local AI

Run the open-weight stack for teams: the mpl host / fleet / node / registry rails for serving, scheduling, and operating local models in production — the path past building a model.

in development

▸ Agent Harness Engineering is the capability layer organizations hire for — see how Mindpool runs the stack for teams →

// the finale

The Research Capstone

An original mini-research study on your own hardware: propose a falsifiable question, run it with the methods the tracks taught, and defend a lab report against a published rubric. Passing work earns on-chain attestation when the marketplace (v2) arrives.

Not everyone passes — by design.

OriginalityMethodologyDefensibilityReproducibility

Requires: finish one specialization track and its capstone first.

On-chain attestation of passing capstones arrives with the marketplace (v2).

See the capstone module →

Ready to start module 01?

Enroll through your mentor to unlock the full curriculum and start building locally.