Time-bounded paths

Build Paths

Pick the path that matches the time you have. Each path is concrete: clone, copy, run. No fluff. By the end of every path you have a deliverable a tutor can review.

30 minutes

Hello-World demo

Students who have never used Azure or Foundry before.

Steps

  1. Clone the sample template: git clone https://github.com/halla-ai/hackathon-sample-2026 && cd hackathon-sample-2026
  2. Run the mock-mode UI: make dev-mock — the UI works without any Azure credentials so you can see what you are building.
  3. Copy .env.example to .env and paste the tutor-provided endpoint and deployment name.
  4. Switch off mock: export USE_MOCK=false. Open the page, send one prompt, observe the model response and the cost.

Token estimate

~5,000 tokens (< 10 KRW model cost)

Deliverable

A screenshot of the first live response.

Common pitfalls

  • Pasting the api-key into the source code instead of .env
  • Not running make dev-mock first — Azure errors hide UI bugs
2 hours

One task working end-to-end

Students who finished the 30-minute path and want a working demo.

Steps

  1. Edit prompts/system.md to one paragraph describing your assistant.
  2. Replace the sample data file (data/sample.md) with 2-3 lines of your real domain content.
  3. Adjust the UI title and the example queries in src/components/SuggestedPrompts.tsx.
  4. Run two test queries — one in-scope, one out-of-scope. Confirm the assistant refuses the out-of-scope query.
  5. Save the two transcripts to demo/transcripts.md.

Token estimate

~30,000-50,000 tokens (< 100 KRW model cost)

Deliverable

A 30-second demo video and transcripts.md.

Common pitfalls

  • Long system prompts (>500 tokens) eat budget on every call
  • Not testing the refusal path — judges always ask
1 day

Technique prototype + polished demo

Teams aiming for the leaderboard or refining the prototype into a second-semester project.

Steps

  1. Pick one of the 6 sample tracks: CV, agent triage, semantic recommender, vision reader, streaming chat, or evaluation harness.
  2. Open the matching src/techniques folder and run its mock/local example before touching Azure.
  3. Replace the sample corpus, rubric, tool, image, or prompt with your own public/synthetic data.
  4. Use only the allowed technology list: gpt-4o-mini chat, JSON, tool calls, streaming, vision input, local vectors, or tutor-enabled embeddings.
  5. Polish the UI: title, empty state, suggested prompts, evidence panel, and result states.
  6. Record a 90-second screen recording: problem → user → demo → impact.
  7. Write the README risk register (5 lines): prompt injection, hallucination, privacy, cost, mitigation.

Token estimate

~200,000-400,000 tokens (roughly 150-400 KRW model cost)

Deliverable

A live working demo, recorded video, 5-slide deck, and README.

Common pitfalls

  • Copying a sample without changing the problem — judges score originality
  • Using excluded Azure services — the demo may fail under the shared plan
  • Forgetting the demo backup — live demos fail at the worst time
  • Treating responsible-AI as optional — it is 10% of judging

Cross-Path Checklist

No matter which path you take, every team should be able to check these six boxes before demo time.

  • sample template cloned and mock mode runs
  • .env populated, never committed
  • system prompt edited for your domain
  • two test queries documented
  • demo backup (screenshot or video) saved
  • README has a 5-line risk register
Follow the deploy walkthrough See sample projects D-Day flow Azure Lab Guide