Semester 4 · Capstone
Capstone Project
https://epochschool.com/subjects/capstone-project
Your capstone is a solo research-and-engineering project in your track (language models or computer vision), run as five graded milestones: a proposal, data and baselines, the main model, evaluation and ablations, and a final report with a reproducibility package, a runnable demo and a viva.
The eight chapters teach the method rather than new models: choosing a falsifiable question that fits your budget, reading and positioning against papers, documenting data and building honest baselines, making every number reproducible, spending a fixed CPU budget wisely, comparing systems with uncertainty, analysing errors and reviewing your own system responsibly, and writing up and defending the work. Everything runs offline on one CPU core with no pretrained weights, so you learn to scale a question to the compute you have.
- Prerequisites
- Required: Finish this subject before you start. The subject builds directly on it.Recommended: Helpful background. You can start without it if you are willing to fill the gaps.
What you’ll be able to do
- LO1Formulate a feasible, falsifiable research question in your track and write a proposal that fixes the primary metric, the minimum effect that matters, the test-set power, the compute budget and a risk plan before any results exist.
- LO2Read papers critically, judge the strength of their experimental evidence, reproduce a reference result at reduced scale, and position your contribution precisely against related work.
- LO3Document data provenance and licences in a datasheet, design splits and leakage checks that match the claim being tested, and build a trivial-to-strong baseline ladder with one evaluation harness.
- LO4Plan and run the main experiments within a fixed one-core compute budget: measure throughput, validate proxies, extrapolate learning curves, tune baselines and the main model fairly, and account for selection bias.
- LO5Design ablations and comparison studies, quantify uncertainty from both training seeds and test items, apply paired tests and multiple-comparison corrections, and report effect sizes with intervals.
- LO6Make results reproducible by construction with canonical configs and hashes, a named seed tree, environment capture, an append-only run log and a package that a verifier and a fresh rerun can check.
- LO7Analyse errors systematically (seeded failure samples, taxonomies, prevalence intervals, slices, behavioural tests) and carry out a responsible-AI review and threats-to-validity analysis of your own system.
- LO8Communicate the project in a research report whose every claim traces to logged evidence, give an honest seeded demo, and defend design decisions and code in a viva.
Chapters
The chapter and lab list for this subject is being finalised and will be published here.
Milestones
- M1The project proposal and a feasibility pilot
- M2Data, splits, leakage checks and the baseline ladder
- M3The main model, a compute plan and fair tuning
- M4Test-set evaluation, ablations, error analysis and the responsible-AI review
- M5The final report, the reproducibility package and the demo
Before you start
Required
Finish this subject before you start. The subject builds directly on it.
Linear Algebra & Calculus for ML
Probability & Statistics for ML
Classical AI: Search, Logic & Reasoning
Data Engineering & Databases for ML
Neural Networks & Deep Learning
Optimisation & Training at Scale
Sequence Models & Transformers
Responsible AI, Ethics & Governance
Recommended
Helpful background. You can start without it if you are willing to fill the gaps.
NLP & Language Model Foundations
Building and Training LLMs from Scratch
Image Formation and Classical Computer Vision
Deep Visual Recognition
How it’s taught
You read short sections written for the page. Each ends with a check, and in-browser exercises let you practise. Graded labs run on your own cloud lab machine, CPU-friendly.
No videos. No live classes. No AI tutor.