AI Hackathon Project Ideas: 30 Buildable Concepts in 72 Hours

If you want to win an AI hackathon, you don't need a "crazy new model." You need a clear problem, a demo that works, and proof it helps.

This list is built for exactly that: 30 ideas you can actually build in 72 hours, grouped by difficulty, and written so you can copy them straight into your team doc and start building immediately.

Before the ideas, one quick rule:

A hackathon idea is only good if you can demo it in 30 seconds. Input → AI → output → action.

The "72h buildable" template#

Every idea below follows the same structure. Use it for any concept, including your own:

• Demo sentence: what the judge sees in one line
• Data needed: what you need to make it real
• Stack suggestion: a safe default stack to ship fast
• Evaluation idea: how you prove it works (tiny but credible)

Easy (10) — high success rate, fast demos, great for beginners#

1. Support Ticket Triage Copilot#

• Demo: Paste a support ticket → it tags topic + urgency → drafts a reply.
• Data: 20–50 sample tickets (real or synthetic).
• Stack: Next.js + FastAPI, LLM classification + template reply.
• Evaluation: Accuracy on 20 labeled tickets + a "time saved" estimate.

2. Meeting Notes → Action Items Extractor#

• Demo: Paste notes/transcript → it outputs structured action items with owners + deadlines.
• Data: 10 meeting transcripts (public or synthetic).
• Stack: React + Node/Express, structured JSON output.
• Evaluation: Precision/recall on action items vs a manual list.

3. CV/Resume Screener (Fair + Transparent)#

• Demo: Upload resume → it extracts skills + matches to role requirements → explains the match.
• Data: 20 resumes (anonymized/synthetic) + 1 job description.
• Stack: Next.js + FastAPI, extraction + scoring rules.
• Evaluation: Consistency tests + an "explanation quality" rubric.

4. PDF Policy Q&A (RAG with Citations)#

• Demo: Upload PDF → ask a question → get an answer with highlighted citations.
• Data: 3–10 PDFs (policies/manuals).
• Stack: Next.js + FastAPI, RAG with vector search.
• Evaluation: Answer correctness on 15 questions + citation grounding rate.

5. Invoice / Receipt Field Extractor#

• Demo: Upload invoice → it extracts vendor, total, VAT, date → exports CSV/JSON.
• Data: 20 invoices/receipts (sample set).
• Stack: FastAPI + simple UI, OCR if needed, extraction.
• Evaluation: Field accuracy per field (e.g., totals correct %).

6. Reply Assistant (Tone + Policy Safe)#

• Demo: Paste message → choose tone → it drafts a reply with the key points included.
• Data: 20 example messages.
• Stack: React + Node, prompt templates + guardrails.
• Evaluation: Human rating checklist (clarity, correctness, tone).

7. Personal Study Coach (Flashcards Generator)#

• Demo: Paste lecture text → it generates flashcards + a quiz → tracks weak areas.
• Data: 3–5 lecture notes.
• Stack: Next.js only, local storage, LLM generation.
• Evaluation: Coverage check (key terms included) + quiz accuracy improvement.

8. Code Explainer + Bug Finder (Small Scope)#

• Demo: Paste code → it explains what it does → flags likely bugs → suggests a fix.
• Data: 10 code snippets with known issues.
• Stack: Next.js + LLM, optional static analysis.
• Evaluation: Bug detection rate on known issues.

9. "Requirements → User Stories" Generator#

• Demo: Paste a feature request → it outputs user stories + acceptance criteria + edge cases.
• Data: 10 feature requests.
• Stack: React + Node, structured output.
• Evaluation: Completeness score vs a checklist.

10. Event/Community Recommendation Bot#

• Demo: Tell it your interests → it recommends events/resources → explains why.
• Data: A curated list of 50 links (manual).
• Stack: Next.js, lightweight search + LLM summarizer.
• Evaluation: Relevance rating from 5 users.

Medium (10) — more integration, more "product feel," still very doable#

11. Incident Log → Root Cause + Next Actions#

• Demo: Paste incident log → it extracts symptoms → proposes root cause + runbook steps.
• Data: 30 incident logs (synthetic ok) + runbook snippets.
• Stack: Next.js + FastAPI, extraction + RAG.
• Evaluation: Correct-action rate on 10 scenarios + citation grounding.

12. Sales Call Analyzer (Objections + Next Steps)#

• Demo: Paste transcript → it highlights objections → suggests a follow-up email + next steps.
• Data: 10 call transcripts.
• Stack: Next.js + Node, structured summarization.
• Evaluation: Objection-detection accuracy + follow-up quality rubric.

13. Knowledge Base Cleaner (Duplicates + Missing Topics)#

• Demo: Upload a doc set → it clusters duplicates → flags missing topics → suggests a restructure.
• Data: 30–100 docs (can be small).
• Stack: Python + embeddings + a simple UI.
• Evaluation: Duplicate-detection precision on labeled pairs.

14. Job Matching for Hackathon Teams#

• Demo: People enter skills → it forms balanced teams → explains team composition.
• Data: 30 fake profiles + skill tags.
• Stack: Next.js + Node, matching algorithm + LLM explanations.
• Evaluation: Balance score (coverage across roles) + user satisfaction.

15. Warehouse/Factory Safety Checklist Assistant#

• Demo: Describe a situation → it outputs a checklist + risk flags + required PPE.
• Data: Safety guideline PDFs.
• Stack: RAG Q&A with citations.
• Evaluation: Correct-citation rate + "missing critical step" rate.

16. Product Feedback Analyzer (Themes + Priorities)#

• Demo: Paste a feedback CSV → it clusters themes → suggests top priorities + rationale.
• Data: 200 feedback lines (public datasets exist).
• Stack: Python + embeddings + a simple dashboard.
• Evaluation: Cluster coherence + top-theme accuracy vs manual.

17. Travel Plan Builder (Constraints + Budget)#

• Demo: Input dates/budget → it creates an itinerary → exports to a calendar format.
• Data: None (use rules + public info placeholders).
• Stack: Next.js + LLM; avoid live web calls in the demo.
• Evaluation: Constraint-satisfaction checks (budget, times).

18. Quiz Generator with Difficulty Control#

• Demo: Upload content → generate a quiz (easy/med/hard) → show answers + reasoning.
• Data: 3 content sources.
• Stack: Next.js only.
• Evaluation: Difficulty calibration (human rating) + coverage.

19. GitHub PR Review Assistant (Rules + Style Guide)#

• Demo: Paste a diff → it reviews against team rules → suggests improvements.
• Data: A style guide + sample diffs.
• Stack: Node/Python + a structured rubric.
• Evaluation: Rule-violation detection rate.

20. AI Form Autofill for Boring Admin Work#

• Demo: Paste an email → it fills a structured form (name, address, request type) → export JSON.
• Data: 50 messages + a target schema.
• Stack: Extraction + schema validation.
• Evaluation: Field accuracy + schema-validity rate.

Hard (10) — bigger "wow," but still feasible if scoped tightly#

21. Multi-Step Research Agent with Verification#

• Demo: Ask a question → it searches your doc set → answers + verifies with sources.
• Data: Curated docs + 20 questions.
• Stack: Next.js + FastAPI, tool-use + RAG, caching.
• Evaluation: Grounded-answer rate + citation correctness.

22. Vision: Defect Detection (Simple)#

• Demo: Upload an image → it marks defect regions → outputs a severity score.
• Data: 50–200 labeled images (or synthetic).
• Stack: Python + a simple CV/vision model, minimal UI.
• Evaluation: Precision/recall on defect detection.

23. "Build a Ticket from Screenshot" Agent#

• Demo: Upload a screenshot → it extracts the issue → drafts a Jira-style ticket with steps.
• Data: 20 screenshots + expected ticket fields.
• Stack: Vision + extraction + template.
• Evaluation: Field accuracy + reproducibility score.

24. Supply Chain Delay Predictor (Toy but Realistic)#

• Demo: Upload a CSV → predict delays → show top risk factors.
• Data: A public logistics dataset or synthetic.
• Stack: Python (scikit-learn) + a simple dashboard.
• Evaluation: Baseline comparison (naive vs model) + MAE.

25. "AI QA Engineer" for a Small Web App#

• Demo: Give a URL + spec → it generates test cases → flags failures.
• Data: A demo app + spec.
• Stack: Node + Playwright + LLM for test generation.
• Evaluation: Number of valid tests + failure-detection rate.

26. Privacy-Safe Local LLM Assistant (Offline Demo Mode)#

• Demo: Ask questions → it answers using local docs → no internet needed.
• Data: Local docs.
• Stack: A local model + lightweight UI, caching.
• Evaluation: Latency + grounded-answer rate.

27. Fraud/Risk Flagging for Transactions#

• Demo: Upload transactions → it flags suspicious patterns → explains the reasons.
• Data: A public fraud dataset or synthetic.
• Stack: Python + anomaly detection + an explanation layer.
• Evaluation: Precision@k on flagged items.

28. "Customer Health" Predictor for SaaS#

• Demo: Upload usage data → predict churn risk → recommend actions.
• Data: A synthetic SaaS usage dataset.
• Stack: A Python model + dashboard.
• Evaluation: ROC-AUC + an action-usefulness rubric.

29. Industrial Maintenance Copilot (Docs + Sensor Notes)#

• Demo: Combine logs + manual → propose a diagnosis + steps with citations.
• Data: Manuals + logs.
• Stack: RAG + extraction, citations, caching.
• Evaluation: Correct-step coverage on 10 cases.

30. AI Demo-to-Deck Generator (Pitch Automation)#

• Demo: Provide a demo summary + screenshots → generate 4-slide pitch deck content.
• Data: 5 sample project summaries.
• Stack: Next.js + LLM + a structured slide schema.
• Evaluation: Slide completeness + clarity score.

How to choose the right idea (fast)#

Pick an idea that satisfies all three:

1. Demoable in 30 seconds
2. Uses data you can get today
3. Has a tiny evaluation you can show

If you're stuck, pick one of these "always works" categories:

• RAG with citations
• Extraction to structured output
• Classification/triage
• Simple vision detection (only if the data is clean)

Read next#

If you want the playbooks that make these ideas win:

• How to Build a Demo in 72 Hours
• Hackathon Tech Stack Guide
• How to Win an AI Hackathon

Bring this to Since AI#

If you want to build with serious teams on real company challenges, bring one of these ideas to the Since AI Hackathon in Turku, Finland. Join the community to find a team and get started.