Lessons that belong to a subject track rather than to one industry programme. Each is a notebook you do, with the same exercises, checks and hints as every other lesson here; the worked solution and full rubric come with enrolment on Synapsa. Where a track has more lessons than exist so far, the rest are planned, not specified in this repository yet.
| Track | Lesson | What you build | Open |
|---|---|---|---|
| T00 · compute | T00-L01-the-8gb-track |
The profiler and the tier gate that decide whether any lesson in this repository may ship. Most programmes start here. | Colab |
| T03 · language models | T03-L01-bpe-from-scratch |
Byte-level byte pair encoding from scratch, and the token tax it imposes on five writing systems. | Colab |
| T03 · language models | T03-L02-bpe-merge-loop-in-cpp |
The BPE merge loop in C++, and its speed-up over the Python trainer on one identical corpus. | Colab |
| T03 · language models | T03-L03-nanolm-cpu-slice |
A character-level language model trained on a CPU with numpy alone, its backward pass derived by hand and checked against finite differences. | Colab |
| T06 · retrieval and RAG | T06-L01-retrieval-from-scratch |
BM25, a dense retriever with no external model, and hybrid fusion, scored by an evaluation harness whose intervals say when one retriever is really better. | Colab |
| T06 · retrieval and RAG | T06-L02-chunking-and-context-budget |
Chunking and context assembly under a token budget, measured by whether the evidence an answer needs actually reaches the model. | Colab |
| T06 · retrieval and RAG | T06-L03-reranking-measured |
A learned reranker, pointwise and pairwise, whose gain is stated with its interval, what reranking depth costs, and the trap of tuning on test queries. | Colab |
| T07 · agents and protocols | T07-L01-agent-loop-from-scratch |
An agent harness with validated tools, budgets, retries and traces, and a measurement of how per-step reliability compounds over long tasks. | Colab |
| T07 · agents and protocols | T07-L02-mcp-from-the-wire-up |
A Model Context Protocol server and client built from the specification, checked by a conformance harness built from the same specification. | Colab |
| T10 · regulation | T10-L01-ai-act-conformity-pack |
An AI system registry turned into a machine-checkable EU AI Act conformity evidence pack. The EU AI Act programme starts here. | Colab |
Each lesson's meta.yaml lists its prerequisites. T00-L01 and T10-L01 need none; beyond them:
- T03-L01 builds on T00-L01.
- T03-L02 builds on T03-L01.
- T03-L03 builds on T00-L01.
- T06-L01 builds on T00-L01.
- T06-L02 builds on T06-L01.
- T06-L03 builds on T06-L01 and T06-L02.
- T07-L01 builds on T00-L01.
- T07-L02 builds on T07-L01.
Synapsa Commons · © 2026 RealAI · licensed under CC BY-NC-SA 4.0. You may use this material to learn and to teach, and copy, fork, share and adapt it. You must credit "Synapsa Commons by RealAI" with a link to https://github.com/TarrySingh/Artificial-Intelligence-Deep-Learning-Machine-Learning-Tutorials, say what you changed, and share adaptations under the same licence. You may not use it, or anything adapted from it, for commercial advantage or payment; for a commercial licence, contact RealAI. The full summary is in NOTICE; the licence's legal code governs.