A certificate is a claim. A live URL is evidence. LevelUpAI is built so every attending student produces something a hiring manager can verify in two minutes.
Every student who attends the required sessions leaves with a public GitHub repository containing a working AI application, a recorded demonstration, and a passed five-minute viva on their own code.
Deployed public URL (or a screen recording of a local run if infrastructure prevents deploy). Open it on a phone.
Code, README with architecture and evaluation table, commit history across multiple days.
Five questions about their code and decisions. Scored against a written rubric. Fakes collapse fast.
| Outcome band | ~Students | What they have |
|---|---|---|
| Advanced | ~15 | Deployed + Edge feature + measured improvement they can explain |
| Strong | ~55 | Deployed public URL, Plus feature, accuracy number in README |
| Basic | ~35 | Working app, recorded demo, passed viva |
| Did not finish | ~15 | Attended some sessions, no complete artifact |
We report ~88% completion honestly, including those who didn't finish. A vendor who reports 100% is not believed — and shouldn't be.
"Built and deployed a document question-answering web application (Python, Streamlit, sentence-transformers, Gemini API) that retrieves relevant passages from user-uploaded PDFs and generates cited answers. Live: [url] · Code: [github]"
"Built a feedback analysis tool now used by the [Dept] office to process ~800 free-text responses per semester: automated cleaning, LLM-based theme classification validated against 30 hand-labelled samples (87% agreement). Code: [github]"
"Built a tool-using retrieval agent; designed a 30-case evaluation harness and improved end-to-end task accuracy from 58% to 71% by adding cross-encoder re-ranking and argument validation. 27 tests, CI on GitHub Actions."
"Built an offline semantic image search tool: indexed 1,200 photos with OpenCLIP, implemented cosine-similarity retrieval and zero-shot classification, measured precision@5 (0.73 → 0.86). Runs entirely on CPU with no internet."
100-point rubric across all tracks. Forty points are automated. The viva only asks about their code.
| Component | Points | How |
|---|---|---|
| Working artifact | 40 | check.py — 8 baseline behaviours |
| Viva | 25 | 5 questions × 5 pts, written rubric |
| Documentation | 15 | README checklist + eval table |
| Process evidence | 10 | Commit history script |
| Demo video | 10 | 90s covering problem / build / bug |
Students memorise and rehearse this structure — it is also the structure of a good interview answer:
Opening viva question, every time: "Show me the part of this you're most proud of, and tell me why."