Insights
What wins
How winners differ from everything submitted. Tags cover 31,547 projects (31,546 by an LLM, the rest by the keyword fallback).
Tag share: winners vs all submissions
Share of projects carrying each tag. A project can have up to 3 tags, so shares don't sum to 100%. Sorted by winner share.
Table
| Tag | Winners | Winner share | All | All share |
|---|---|---|---|---|
| agents | 224 | 38.3% | 10772 | 34.1% |
| productivity | 138 | 23.6% | 9168 | 29.1% |
| vision | 122 | 20.9% | 3254 | 10.3% |
| voice | 114 | 19.5% | 3076 | 9.8% |
| dev tools | 103 | 17.6% | 5856 | 18.6% |
| health | 77 | 13.2% | 2750 | 8.7% |
| education | 63 | 10.8% | 4120 | 13.1% |
| creative/music | 63 | 10.8% | 3024 | 9.6% |
| science | 52 | 8.9% | 2271 | 7.2% |
| finance | 37 | 6.3% | 2549 | 8.1% |
| robotics/hardware | 30 | 5.1% | 787 | 2.5% |
| security | 29 | 5% | 2716 | 8.6% |
| accessibility | 29 | 5% | 917 | 2.9% |
| legal | 23 | 3.9% | 1194 | 3.8% |
| culture/heritage | 12 | 2.1% | 467 | 1.5% |
| repair/trades | 7 | 1.2% | 326 | 1% |
Which labs reward which tags
Lift = a tag's share among a lab's winners ÷ its share among that lab's submissions. Blue: winners over-index; red: under-index; gray: no difference. Blank: fewer than 3 winners or no submissions with the tag.
| agents | productivity | vision | voice | dev tools | health | education | creative/music | science | finance | robotics/hardware | security | accessibility | legal | culture/heritage | repair/trades | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Google 319 / 17,125 | 1.7 | 2.2 | 1.8 | |||||||||||||
OpenAI 92 / 11,718 | 2.6 | 3.9 | 3.1 | 1.9 | ||||||||||||
Anthropic 113 / 1,540 | 2.5 | 1.7 | 1.9 | 2.0 | 1.5 | 5.8 | ||||||||||
Hugging Face 0 / 572 | ||||||||||||||||
ElevenLabs 23 / 303 | 2.5 | 1.8 | 3.3 | |||||||||||||
Meta 27 / 182 | 2.1 | 3.4 | 1.7 | 2.0 | 1.9 | |||||||||||
NVIDIA 4 / 71 | 1.7 | 3.5 | ||||||||||||||
Mistral 7 / 36 | 1.7 | 5.1 | 2.6 |
Tag share over time
Top 6 tags by winner share, by the quarter each hackathon started. Quarters with fewer than 10 projects are hidden.
Most common core tricks among winners
Two-word phrases that recur across winners' core tricks.
- BEAMTEAM · Gemini 4 Argon SF Hackathon
Parallel agents extract a layout from a paper figure, then a deterministic beam tracer checks power, polarization and laser safety before export.
- Morf · Gemini 4 Argon SF Hackathon
Generates CAD assemblies from text, renders the geometry, then runs dimension and motion checks to verify and correct each result.
- Aniket's Team · Gemini 4 Argon SF Hackathon
Fans out 20+ managed agents, one per studio role (3D, level, music, sound, gameplay), to produce a playable 3D game in hours.
- polar bears · Gemini 4 Argon SF Hackathon
Builds a MuJoCo bench twin from a photo, simulates the protocol stepwise, and traces each failed well to its causal step.
- Eagle · Gemini 4 Argon SF Hackathon
Code computes VFR risk and fuel limits, while a cloud agent re-plans in flight and texts one decision over a 160-character link.
- Revive · Gemini 4 Argon SF Hackathon
Gemini watches all surviving videos in one long-context call, then a builder agent rebuilds the game and checks each rule against timestamped footage.
- NurseMate · Intelligence at Work Hackathon
Phone agent transcribes both call sides live and stores every extracted fact with a verbatim quote in Snowflake.
- HealthForce · Intelligence at Work Hackathon
Cortex Agent ranks candidate patient identities from partial descriptors, then merges records across care sites and parses spoken orders into structured task changes.
Top starred repos
GitHub stars at last enrich. Orange: won an award.