Learn It, Then Prove It

Certification

Courses built the way we work: every module ends in something that runs. Finish a track, pass the project review, and the certificate points at a repo — not an attendance record.

The Shape Of It

How It Works

Three stages per track. You can stop after the first one and still walk away with something — the certificate is the last step, not the point of the whole thing.

01

Work Through the Modules

Each module is written material plus an interactive artifact and a small exercise. No video lectures to sit through — you read, you run, you break something, you fix it.

02

Ship the Capstone

Every track ends with one build brief and a set of acceptance criteria. You put it in a public repo. It has to actually run on ordinary hardware — that's the bar.

03

Review & Certify

We review the repo against the criteria and send written feedback either way. Pass, and you get a certificate with a verification ID that links to the work itself.

The Catalogue

Courses

Six tracks. Take them in order for a full path from "I run a model locally" to "I run a system", or take just the one you need.

Track 01 Enrolling Soon

Local-First AI Foundations

Get a real model running on your own machine and understand what you're actually looking at — weights, quantization, context windows, and where the memory goes.

Beginner 5 modules ~12 hours
  • Run and serve local models with Ollama and llama.cpp
  • Read a model card and predict whether it fits your VRAM
  • Choose a quantization level on evidence, not vibes

Capstone A local assistant with a documented hardware profile and measured tokens/sec.

Track 02 Enrolling Soon

Prompt Engineering for Production

The part everyone skips. Prompts as versioned artifacts with evaluation suites behind them — not strings someone edited in a hurry at 2am.

Beginner 4 modules ~10 hours
  • Structure prompts that hold up across model swaps
  • Build an eval set and catch regressions before users do
  • Constrain output formats so downstream code can trust them

Capstone A versioned prompt suite with an automated eval harness and a regression report.

Track 03 Curriculum Ready

Retrieval Systems That Actually Work

Most RAG demos fall over on real documents. This track is about the unglamorous parts: chunking, hybrid search, reranking, and knowing when retrieval is the wrong tool.

Intermediate 6 modules ~18 hours
  • Design chunking strategies that survive real documents
  • Combine dense and keyword retrieval, then rerank
  • Measure retrieval quality separately from answer quality

Capstone A retrieval system over your own corpus with a measured recall@k baseline.

Track 04 Curriculum Ready

Fine-Tuning & LoRA Adapters

When prompting stops paying off, you adapt the model. Dataset construction, LoRA and QLoRA, and how to tell a real improvement from a memorised eval set.

Intermediate 6 modules ~20 hours
  • Build and clean an instruction dataset that teaches one thing well
  • Train a LoRA adapter on a single consumer GPU
  • Evaluate honestly — held-out sets, contamination checks, ablations

Capstone A trained adapter with an evaluation card showing where it wins and where it doesn't.

Track 05 In Development

Agent Orchestration

Tool use, routing, and multi-step planning — plus the failure modes nobody warns you about: runaway loops, silent tool errors, and context that quietly falls off the end.

Advanced 6 modules ~22 hours
  • Design tool interfaces a model can use without hand-holding
  • Route between small and large models on cost and risk
  • Build the guardrails: budgets, timeouts, and recovery paths

Capstone A multi-tool agent with a trace log and a documented failure-and-recovery analysis.

Track 06 In Development

Deployment & Benchmarking

Getting a system off your laptop without handing it to a cloud vendor. Serving, batching, telemetry, and benchmarking your hardware honestly.

Advanced 5 modules ~16 hours
  • Serve models on-prem with sane batching and memory limits
  • Instrument latency percentiles, not just averages
  • Benchmark real hardware against vendor claims and explain the gap

Capstone A deployed service with a reproducible benchmark report and a p99 latency budget.

Enrolment isn't open yet. Curricula and capstone briefs are written; we're finishing the module artifacts before letting anyone pay for anything. If you want a seat in the first cohort — or you're a team wanting a track run privately — get in touch and we'll put you on the list.

What You Walk Away With

The Certificate

What it actually says

A Hallucinated Lab Certified credential states that a specific person built a specific thing that met a published set of criteria, reviewed by a named person on a given date. It carries a verification ID and links to the capstone repository. That's it — no scores, no percentile, no badge inflation.

What a capstone has to clear

  • It runs on ordinary consumer hardware, from a clean clone
  • The README lets a stranger reproduce your results
  • Claims are backed by measurements you took, not numbers you read
  • Limitations and failure cases are documented, not hidden
  • No cloud dependency the brief didn't ask for
  • The work is yours, and AI assistance is disclosed where used

Want a seat in the first cohort, or a track delivered to your team?

Talk to us about enrolment