"Everything you've learned comes together here: pick a question, test it fairly, and share what you find. This is how you do research, not just read it."
Level: Intermediate · Time: ~2–3 weeks · Prerequisites: the whole track
Learning Objectives
By the end of this module, you will be able to:
- Scope a small, doable research project
- Apply the full research loop end to end
- Present findings honestly and clearly
- Turn a project into a portfolio piece
1. Why a Capstone?
Reading about research isn't the same as doing it. A small capstone — one question you investigate properly — cements every skill from this track and becomes proof you can think like a researcher.
Concept: A great project is small in scope but complete: a clear question, a fair test, honest results, and a clear write-up. Better a tidy mini-study than an ambitious one you never finish.
2. Pick a Good Question
Choose something specific and testable that you can run with modest resources.
| Project type | Example |
|---|---|
| Reproduction | Reproduce a small result from a paper (Module 4) |
| Ablation | Does component X actually help? (Module 5) |
| Comparison | Which of two methods wins on a fair test? |
| Application | Apply a known method to a new dataset |
A concrete, right-sized example: "Does adding data augmentation improve a small image classifier on a 5,000-image dataset?" It has a clear question, a fair comparison (with vs without augmentation), a metric (accuracy over 3 seeds), and it runs on a free GPU in an afternoon. That single, finishable study demonstrates every skill in this track far better than an abandoned attempt to "build a better language model."
Common mistake: Picking a question so big it needs a lab and a month of GPU time. Scope down until you can get a first result in days — you can always expand later.
3. Run the Research Loop
Apply what you've learned, step by step:
- Review — find the key related papers (Modules 2–3).
- Hypothesize — state a testable claim (Module 5).
- Build — reproduce or implement the method (Module 4).
- Experiment — fair baselines, controls, multiple seeds (Module 5).
- Evaluate — sensible metrics, honest reporting (Module 16).
- Write up — question, method, results, limitations.
Try this: Before running anything, write your hypothesis and what result would prove you wrong. Pre-committing keeps you honest when the data comes in.
4. Share It Well
Put your project on GitHub with a clear README: the question, what you did, your results (including plots), and honest limitations. Add the code so others can check it. A small, well-documented, reproducible study speaks louder than a flashy claim.
A strong README follows the same skeleton as a paper: What question did I ask? What did I do? What did I find (with a plot or table)? What are the limitations? Add a one-line "how to run" so a stranger can reproduce your result in minutes. This structure isn't busywork — it's the exact reading pattern you learned in Module 2, now applied to your own work so others can judge it fairly.
Concept: Honesty is your credibility. Reporting "X didn't help as much as expected" is a real, valuable finding — and shows the rigor that defines good research.
5. From Project to Path
Your capstone is a launchpad: it can grow into a blog post, a contribution to an open-source project, or the seed of a bigger study. It also tells employers and collaborators that you can investigate, not just recite.
Real-world use case: Plenty of people have turned exactly this kind of small, honest study into their first real footprint in the field — a reproduction that surfaced a subtle bug, a clean ablation that others cited, a write-up that landed an interview. The value isn't the size of the result; it's the visible proof that you can run the research loop end to end and report it truthfully.
✅ Checkpoint
- What makes a great capstone (per the key idea)?
- List the six steps of the research loop.
- What should a project's write-up always include?
Answers: 1) Small in scope but complete — clear question, fair test, honest results, clear write-up. 2) Review, hypothesize, build, experiment, evaluate, write up. 3) The question, method, results (with limitations), and code to reproduce it.
Key Takeaway: The capstone is where you prove you can do research: pick a small but complete question, run the full loop — review, hypothesize, build, experiment, evaluate, write up — and share it honestly on GitHub with code and limitations. Completeness and honesty beat ambition; a tidy, reproducible mini-study is a genuine portfolio piece.
Further Learning
- Papers with Code — find reproducible baselines
- Distill.pub — models of clear research communication
Part of "Research & Advanced AI." Original content for this learning platform.