Blog · Nov 16, 2026
Drexathon recap: what builders made with Drex in one day
Nace.AI
Drex, a small model that decides
Generally available. Ahead of Jev on the public Decision Index 0.3.1, with about 3B parameters active per token.

On Saturday, 14 November, we hosted Drexathon, our first hackathon, at Frontier Tower in San Francisco, with Adobe EDU as our partner. Teams of one or two spent the day building agents that make decisions with Drex, our decision model. They read documents with Nace Document Intelligence (NDI) and ran on Intern, our open-source agent harness.
Drexathon in numbers
- 355 people registered on Luma.
- About 150 builders came on the day.
- 3 tracks, plus a Best Use of Adobe prize open to every team.
- 4 hours of building, from 11:00 AM to the 3:00 PM submission deadline.
- More than 45 projects submitted as pull requests to the public Intern repository.
- 5 minutes for each finalist demo.
How the day ran
| Time | What |
|---|---|
| 9:30 AM | Check-in and breakfast |
| 10:00 AM | Welcome and kickoff |
| 10:15 AM | Product demos from Nace.AI, rules and judges |
| 11:00 AM | Building starts |
| 3:00 PM | Hard submission deadline |
| 3:15 PM | Finalist demos |
| 4:30 PM | Awards |
| 4:45 PM | Networking |
Teams built on Intern and submitted their code as pull requests to the Intern repository, so the projects are public.

The tracks
🦖 Apex Predator: Drex vs. the Giants. Take a task a frontier LLM normally handles, give the decisions to Drex, and prove it's faster, cheaper or more accurate. Teams were judged on measured results and the size of the improvement.
🧬 Swarm Intelligence. Build a system where several agents coordinate, compete or negotiate, with Drex making every agent's calls.
🎨 Wildcard. The most creative, surprising or delightful use of Drex.
In every track, teams that used NDI to bring real-world documents into their build earned bonus points. Adobe EDU added a Best Use of Adobe prize for teams that used Adobe tools in their project or demo.
The judges
- Shakh Sailaukan, Co-founder of Rixt.AI
- Nischay Dhankhar, Chief of AI at Nace.AI
- Mark Illes-Toth, Senior PM at IBM
- Kostas Bimpikis, Professor at Stanford
- Aidana Daulbayeva, Product and Strategy at Sierra
The winners

Apex Predator: Drex Vertical Reframe, by team Wukong
Drex Vertical Reframe turns long landscape video into vertical 9:16 shorts. NDI and crowd noise shortlist the moments worth keeping, Drex classifies them, and then Drex decides frame by frame which part of the picture survives the crop. The team turned framing into multiple choice: a numbered grid on each frame, and Drex picks the lines that mark where the action is. In the team's words, every judgment call is a Drex call and everything else is deterministic code. On a 30-minute football match, Drex found all 4 key moments on time, against 1 for Gemini watching the full video. On the same clips, Drex framed each one 4 to 6 times faster and about 20 times cheaper than Gemini, the team reports.
Swarm Intelligence: Drex Poker Arena
Five AI poker players play Texas Hold'em against each other or against you, and every move is a live Drex call. Drex returns a probability for each legal move, and each bot's personality decides what to do with those odds. The Statistician always takes Drex's top move, the Conservative raises only when Drex is more than 60% sure, and the Manipulator asks Drex whether its opponents will fold. After each hand you can see every bot's cards, Drex's probabilities and the reason for each move. In evolution mode, the worst bot goes extinct and the best one spawns a mutated child. NDI reads messy hand-history logs and feeds opponent stats to the bots, and the bot avatars were made with Adobe Firefly. The team reports 368 ms per decision and no illegal moves in its benchmark, against 4 for a rule-based bot.
Wildcard: Reach
Reach is a Mac desktop assistant that acts for you. You can type to it, talk to it or circle something on your screen. Every request is an Intern session, with Drex Max as System 1 and Qwen3-30B-A3B on DeepInfra as System 2. Drex sizes up each request before it runs, including what kind of task it is and how risky, in about 260 ms. It also decides whether speech it overhears is meant for Reach and routes voice commands. NDI finds evidence in documents and returns cited quotes, grounded to the page.
Best Use of Adobe: Boneify
Fossil or fiction? You upload a document and paste what an AI wrote about it. NDI reads the original and pulls out key facts with citations. Drex then labels every sentence of the AI's text as correct, wrong or made up, with a confidence for each. The receipts appear in the original PDF through the Adobe PDF Embed API: a green box for a correct sentence, a red box for a wrong one. Intern can then rewrite only the flagged sentences, and Boneify checks the rewrite again. On 450 answers from the RAGTruth test set, the team reports that Drex Max scored a higher sentence-level F1 than GPT-6.1 Sol, 0.56 against 0.51. It was also 12.9 times faster and 45 times cheaper.
What else people built
The tracks left a lot of room, and the other entries range widely. Teams built:
- a mock jury of 12 where Drex makes every juror's decision,
- an election simulator where Drex casts every vote,
- a real-time capture-the-flag game where Drex commands every unit against an army run by an LLM,
- a simulated robotics match where Drex makes every robot's call,
- a composer that turns a prompt into music, with Drex making every musical decision,
- an allergen checker that reads food-label photos,
- a pipeline that turns driver-camera crash video into an ER pre-alert,
- a tool that rebuilds torn and shredded pages from a photo,
- an inbound firewall that screens every tool result an agent reads,
- and a CI helper that decides whether a failing test is flaky or real.

Partners

- Adobe EDU partnered with us on the event and sponsored the Best Use of Adobe prize. Winning projects used Adobe Firefly and the Adobe PDF Embed API.
- DeepInfra served open models for builders. Reach and Boneify both ran Intern's System 2 model on DeepInfra.
Thanks to Frontier Tower for hosting us, to the judges, and to everyone who came to build.
What's next
The submitted projects are public in the Intern repository, so you can read the code and run it yourself. If you want to build with Drex, you can try it for free in the Nace Console. To hear about the next hackathon, follow our events page.
