Build one portfolio project worth discussing

My preferred first capstone is a document assistant for a narrow user group, with verifiable answers and a measured improvement over search. Use a domain you understand. A small, honest project with real feedback gives you more to discuss than several copied demos.

Choose one problem

• Default: a support knowledge assistant for an open-source tool. It retrieves versioned documentation, cites sources, and says when it cannot answer.
• Operations alternative: a synthetic invoice or ticket intake assistant that extracts fields for human review. Compare against rules and manual completion time.
• Engineer extension: add one read-only integration, a replayable evaluation suite and a deployment failure drill.
• Vibe-coder extension: improve an existing demo’s tests, data flow, error handling and reproducibility before adding another feature.

Five small milestones

1. Problem: speak with two potential users, or state clearly that your user assumptions are unvalidated. Write the brief and success criterion.
2. Baseline: implement search or rules, and measure it on a fixed development set.
3. AI version: add one model-powered step with traceable inputs and outputs.
4. Evidence: evaluate on held-out cases, list failures and observe someone using the tool.
5. Delivery: produce a live or recorded demo, a runnable repository and a short case study.

The finished evidence package

• A 2–3 minute demo showing the ordinary workflow and one failure.
• A README that a reviewer can run, with no embedded keys.
• A clear architecture diagram and data provenance notes.
• An evaluation table with baseline, improved version, sample size, latency and estimated cost.
• A failure log, tradeoff discussion and a concrete next improvement.

Do not invent customers, revenue, production scale or time savings. If the test uses synthetic data, label it. If three friends tested the app, say three testers. Put placeholders in your resume until you have measurements.