Capstone Project Notes
Subject Overview
The capstone is not coursework. It is a real product, built end-to-end, that solves a real problem for real users, applying everything from the previous three years — the focus is entirely on building. It must use Python and at least one ML or LLM model, be deployed and publicly accessible (not a localhost demo), include quantitative evaluation with documented results, be built with AI assistance (Cursor, Claude Code or equivalent, with the vibe coding workflow itself part of the deliverable), and reach at least five real users whose feedback drives at least one iteration.
Unit-wise Syllabus
8 units — click WhatsApp below to get the full notes for each
Unit 1: Discovery & Spec (4–5 hours)
Pick the problem, talk to potential users, write a one-page spec, sketch the architecture
Unit 2: Data & Model (10–12 hours)
Collect or generate data, train/fine-tune a model OR design the LLM pipeline, run baseline evaluations
Unit 3: Build the Product (15–18 hours)
Frontend (Next.js/React), backend (FastAPI), database, auth; using Cursor + Claude Code as your pair programmer
Unit 4: Deploy & Monitor (5–6 hours)
Deploy to production, set up observability and error tracking, add an evaluation harness to CI
Unit 5: User Testing & Iteration (4–6 hours)
Onboard real users, collect feedback, ship at least one substantial improvement
Unit 6: Demo Day (2–3 hours)
Write-up, demo video, public launch (Product Hunt / X / LinkedIn), open-source the relevant parts
Unit 7: Example capstone ideas
Domain-specific copilot — a coding assistant for a niche framework or a writing assistant for a specific industry; RAG-powered knowledge product — a searchable expert system for legal docs, medical papers or university lectures; vision application — a fine-tuned CV model for retail shelf analysis, agriculture or accessibility; AI agent for a real workflow — handling support tickets, generating marketing campaigns or automating a research task end-to-end; creative tool — an AI-powered video editor, music generator or design assistant with a polished UI
Unit 8: Evaluation criteria
Does it work — real users can use it without hand-holding; is the AI integration thoughtful — not just an OpenAI wrapper, with evidence of evaluation, prompt engineering or model adaptation; is the code shippable — tests, basic CI, error handling, observability; did vibe coding accelerate you — a short write-up on how AI tools helped or hindered the build; did you learn from real users — documented feedback and at least one iteration based on it