1
Profile input
Student enters skills, interests, and preferred project style.
All samples · Embeddings
Skill-Matched Recommender
A local-vector scaffold that can later switch to Azure embeddings without requiring Azure AI Search.
Problem
Students describe skills and interests in different words than role descriptions, so exact filters miss good matches.
Users
Students choosing hackathon roles, project ideas, internships, or next learning modules.
Why this track
This practices embeddings, similarity search, and recommendation explanation while respecting the current Azure rule: no Azure AI Search dependency by default.
Architecture
Stay minimal. 5-6 nodes. Each arrow is one network hop.
profile
Student profile text
embed
Embedding model or local vectorizer
corpus
Role/project corpus
ranker
Cosine similarity ranker
evidence
Matched skills + gaps
ui
Recommendation UI
Edges
- profile embed — query vector
- corpus embed — cached vectors
- embed ranker — vectors
- ranker evidence — top-k matches
- evidence ui — explainable result
Prompt Pack
Starting prompts. Iterate. Move the system prompt into prompts/system.md so it can be versioned.
system
You explain recommender results. Use only the matched role text and skill evidence. Return a short recommendation, two skill gaps, and one next learning action. Do not rank people by sensitive attributes. user
Profile: Python, Docker, FastAPI. Top matches: backend internship, MLOps assistant, data analyst.Code Snippet
The pattern shape. Read it, run the matching scaffold, then adapt the idea for your own team.
python
query_vec = embed("Python Docker FastAPI internship")
role_vecs = [embed(text) for text in ROLE_DESCRIPTIONS]
scores = [
(cosine(query_vec, role_vec), role)
for role_vec, role in zip(role_vecs, ROLE_DESCRIPTIONS)
]
top_matches = sorted(scores, reverse=True)[:3]
# ... your turn: explain why each match helps the student
Reference: src/techniques/embeddings_search/ in halla-ai/hackathon-sample-2026
Demo Screens
Three screens that prove the prototype works.
2
Top matches
Three roles or ideas with scores and matched evidence.
3
Learning gaps
Short next-step checklist based on missing skills.
Azure budget
Local vectors are free. If tutors enable text-embedding-3-small, 1M embedding tokens is roughly USD 0.022 before storage.
Pitfalls
- • Symptom: every result looks similar. Cause: corpus items are too generic. Fix: make role descriptions specific.
- • Symptom: rankings are hard to defend. Cause: only score is shown. Fix: show matched terms or evidence.
- • Symptom: cost spikes. Cause: corpus is re-embedded every request. Fix: cache vectors.
Possible Extensions
If you finish the 1-day path early, use one question below to make the project more original.
- How would your version combine semantic score with explicit skill tags?
- How would you show enough evidence that users trust the match?
- How would you update cached vectors when roles change?