✦
Challenge 01
Best Use of Gemma 4
Build with Gemma 4, Google's lightweight, open-weight model, through the
Gemini API. Make its AI capabilities a meaningful part of your prototype.
Useful projects you can prototype in a day
- Form Rescue: A web app that turns confusing form instructions into a guided
checklist. Use Gemma to explain each field and flag missing information without inventing
answers. One-day build: support one form and show completion progress.
- Promise Tracker: Find commitments buried in a pasted conversation: who promised
what, by when, and what is still unresolved. One-day build: use Gemma to create an editable
board with a source quote for every task.
- Repair Before Replace: Describe a broken everyday item and the symptoms; Gemma
asks clarifying questions and retrieves relevant steps from a supplied repair guide. One-day
build: cover bicycle punctures and show the source for each step.
- Buy Together: Help a community combine small purchase requests into one order.
Use Gemma to turn messages like “need two notebooks” into structured items, then
group matching requests in code. One-day build: an editable shared order sheet with quantities
and totals.
Identify the Gemma model in your README
and show its integration in your code or demo.
Explore Gemma
resources →
✦
Challenge 02
Best Open-Source AI Project
Create an original project with open-source or open-weight AI at its core: a
reusable agent skill, an application using a small or large language model, or software that manages
a model's tools and workflow.
Useful projects you can prototype in a day
- Token Budget Lab: A web app that uses an open model to compress long prompts
while preserving instructions and key facts. One-day build: compare original and compressed
token counts for one target model, test both on the same questions, and show compression
overhead alongside potential savings.
- Context Picker: Stop sending an entire document to answer one question. Use an
open embedding model to select relevant passages within a token budget. One-day build: upload
one text document, ask a question, and inspect the selected excerpts and token count before
copying the prompt.
- Safe Share: Preview what sensitive information a support ticket or log would
expose before sharing it with AI. Combine pattern checks with an open model to suggest
redactions. One-day build: paste text, review highlighted findings, and export an approved copy.
- Chat Handoff: Turn a long AI conversation into a compact brief for the next
session. Use an open model to preserve decisions, constraints, failed attempts, and next steps
with source references. One-day build: paste a transcript, edit the brief, and compare token
counts.
Keep the first version to one input, one
useful output, and a few realistic examples. Add accounts, extra integrations, and automation after
hack day.
Requirements
- Use open-source or open-weight AI as a substantial part of the project.
- Publish your code in a public GitHub repository with an open-source license.
- Agent skills must follow the Agent Skill Open Standard.
- Model harnesses must be original implementations or meaningful improvements to existing
open-source harnesses.
More Project Ideas
Build with Limited LLM Access
Useful apps can start with short prompts and small outputs. Let code handle sorting, parsing, and
chunking, then use your model for the decisions that need language understanding. Adapt these ideas to
your chosen challenge and keep its AI requirements in mind.
- Downloads Organizer: Sort files by extension and filename rules first; ask the
model to classify only unknown filenames. One-day build: preview suggested folders, require approval
before moving files, and keep an undo log. Start with filenames rather than file contents.
- Daily Standup Writer: Read a small batch of recent Git commits and draft three
lines: completed work, suggested next steps, and blockers to confirm. One-day build: select a
repository and date range, then edit and copy the draft. Do not invent blockers or treat suggested
plans as facts.
- Receipt & Bill Extractor: Use local OCR or PDF text extraction to read a
receipt; ask the model only to map the text to date, vendor, and amount. One-day build: review
extracted fields and export CSV, leaving uncertain values for correction.
- Notes to Flashcards: Split notes into short chunks in code and ask the model for
one question and answer per chunk. One-day build: paste notes, review cards with their source text,
and export them for revision.
- Email or Chat Triage: Use keyword scores to sort obvious cases and ask the model to
label only ambiguous messages as “reply now”, “later”, or
“ignore”. One-day build: paste a sample inbox and review an editable priority queue
without sending or deleting anything.
- Meeting Action-Item Extractor: Start from a text transcript and extract only who
does what by when, with a supporting quote. One-day build: paste a short transcript, correct the
task table, and export it; flag missing owners or dates. No audio processing needed.
- Voice Khata / Expense Log: Turn a short expense entry into amount, category, and
date, then calculate totals in code. One-day build: start with typed entries and an editable ledger;
add local speech-to-text only if the laptop can run it comfortably. Keep typed input as a fallback.
- Paper Explorer Upload a research paper, ask questions about its methodology,
results, tables, and figures, and get answers grounded in the relevant sections.
- Prompt Optimizer / Token Saver Take a long prompt, remove redundant content,
generate a shorter version while preserving important instructions, and show the approximate token
savings.
- Project Planner / AI AssistantGive the assistant a college-project idea and
generate requirements, features, development tasks, basic architecture, database entities, and a
testing checklist.
Keep it small: Limit input length, process one item at a time, and
let users review model output. Test your chosen model on a few real examples before the event; short
prompts still need accuracy checks.
Before the Event
Download Your Models Before You Arrive
If your project uses a local open-source or open-weight model, including Gemma, download it onto your
laptop before the event. It is best to complete all large downloads at home so you can start building
when you arrive.
- Choose a model that fits your laptop's RAM, available storage, and GPU, if required.
- Complete any account setup, license acceptance, or access requests ahead of time.
- Download the model weights, tokenizer, and required files; install your runtime and project
dependencies.
- Run a sample prompt locally and check that it works without downloading additional files.
- If you use a hosted API, set up and test your API access beforehand and keep a backup internet
connection ready. Follow your chosen challenge's integration requirements.
Download resources: Gemma setup & download guide · Google models on Hugging
Face · Browse open
models · Hugging Face download instructions
Wi-Fi: Wi-Fi will be provided, but reliable connectivity is not
guaranteed. Speeds may vary depending on how many participants are using it at the same time. Please
avoid relying on event Wi-Fi for model downloads and bring a mobile hotspot as a backup.
