L5 · 07 · Capstone: Publish a Speciality Project
What separates "mastered the material" from "is a master" is producing something others can use. The Level 5 capstone is a speciality project you actually publish — a blog post, an open-source tool, a reproduced result, a talk, or a paper. You pick the speciality; the output must be public and useful to someone who isn't you.
The deliverable
- One public artefact in your chosen speciality: blog post, open-source project, reproduced result, talk, or paper.
- Deep enough that it would help a peer — not a summary of someone else's work.
- Honest about limits: what worked, what didn't, what you'd do next.
Pick a speciality
Choose the one that matches where you want to go deep. Each has a suggested first project; the electives on the roadmap give more options.
| Speciality | Example capstone project |
|---|---|
| Applied LLM systems | A production-grade agent or RAG system, open-sourced with evals + a write-up of the design trade-offs |
| Interpretability | A probing/steering study on a small model: "can you find the concept of X in model Y?" |
| On-device / edge AI | A privacy-first feature that runs fully on-device, with the hybrid router benchmarked |
| Multimodal | A document-understanding pipeline with an eval set and honest failure-mode analysis |
| Developer tooling | A useful MCP server, skill pack or IDE workflow you publish for others (e.g. Scala-specific) |
| Research reproduction | A faithful small-scale reproduction of a known result, with ablations and a write-up |
| AI adoption / leadership | A measured pilot in a real team: baseline, intervention, honest before/after |
The shape of a good capstone
- A real problem, ideally one you or someone near you actually has. Motivation matters.
- Depth over breadth. One question answered well beats ten skimmed.
- Evidence. Numbers, evals, ablations — not just "it worked". The discipline from L3·04 and L5·01.
- Public and reproducible. Code, data (or a way to get it), and enough write-up for someone to repeat it.
- Honest. Say what didn't work. A negative result with good methodology is a real contribution.
Formats
- Blog post / write-up: the default. Claim, method, evidence, verdict.
- Open-source repo: a tool with a README someone can run.
- Talk: internal meetup, conference, or a recorded walkthrough.
- Reproduction report: "I re-ran paper X; here's what held up."
- Short paper: if you found something genuinely new, a workshop paper or a solid technical report.
Acceptance criteria
- It's published: a public URL, repo, or talk someone else can access.
- It contains evidence: measurements, evals or a reproduction — not just opinion.
- A peer could use or reproduce it from what you published.
- It states limitations honestly.
- You can explain every design decision in it.
The real graduation
A curriculum ends; mastery doesn't. The capstone is the proof you can generate new value in this field, not just absorb it.
Publish it, get feedback, and let it point at the next thing you go deep on.
Where to go from here
Follow the electives, pick a research area that excites you, contribute to open-source, mentor someone earlier in the curve
— or write the docs you wish you'd had. Teaching is the last and best way to master a subject.