Two Years of a Knowledge Repo: What I Actually Learned
A data-driven retrospective on ~600 notebooks across 18 repos since Sep 2024 - how the work shifted from foundations to LLMs to real APIs to hardware.
I started the Knowledge mono-repo on 1 September 2024. Almost two years and ~600 notebooks later, the git history tells a clear story about how my focus moved. This is that story, from the data.
The arc, in one table
| Period | Theme | What landed |
|---|---|---|
| 2024 Q3-Q4 | Foundation | Django/backend, front-end, Linux/dev tooling, Docker |
| 2025 H1 | Depth | First LLMs, math/physics notes, AWS, stock/finance |
| 2025 H2 | Real-world | Client APIs (Cliniko/Nookal/Snapforms), AWS deep dive |
| 2026 | AI systems + hardware | RAG course, LLM pipelines/agents, IaC, motors/ESCs |
Notebooks added per quarter
The 2024-Q4 spike is the initial build-out - most of the reference library was seeded at once. After that, steady growth with a clear AI-and-hardware resurgence in 2026.
Where the work went (commits by repo)
Finances, Other, WEB_doc, and Back_End dominate - the repos tied to real projects and daily use, not just reference.
Topic evolution
flowchart LR F[Foundations 2024] --> L[LLMs 2025] L --> A[Real APIs 2025-H2] A --> R[RAG + agents 2026] R --> H[Hardware / embedded 2026]
Four things I actually learned
- The biggest AI value came late and fast. The RAG course and LLM pipelines (2026) are the single largest coherent build-out - see Building a production RAG system.
- Reference notes pay off when they meet a real project. The repos that grew most (Other, WEB_doc, Finances) are the ones wired to actual work.
- Tooling maturity compounds. Moving to uv/Ruff/Polars made every later notebook cheaper to write.
- Breadth then depth. 2024 was breadth (seed everything); 2026 is depth (go deep on RAG, on motors, on IaC).
The site that holds it all together is itself a project - how the federated docs site works.