Your project is judged live, on the day, in front of the panel. These rules and the same 100-point rubric apply to all four tracks.
First, pass the live-run check.
This is an eligibility gate, checked before scoring. It does not carry points.
Counts
A build the judges can operate themselves, running on real input, end to end.
Does not count
Slides. Figma prototypes. Click-through mockups. Pre-recorded video in place of a live run. A demo that only works offline on your machine with data you typed in beforehand.
The judges will ask:
Can I operate the build myself?
Can I try an input I choose right now?
Can I run it twice and get consistent results?
If it cannot run in the room, it does not advance.
100 points. Four criteria.
Technical execution carries the most weight because the event is about building something that works in 24 hours. Track-specific topics are assessed within this rubric, without extra or hidden weights.
Technical execution
40pts
Does it actually run, and how deep is the build underneath?
Usefulness
25pts
Does someone genuinely need this?
Originality
20pts
A new approach, or a reskinned tutorial?
Demo quality
15pts
Three minutes. Is the value obvious?
Score each criterion from 1 to 5.
5
Exceptional: worth taking forward after the event.
4
Strong: runs end to end with polished execution.
3
Adequate: works, but lacks depth or has significant gaps.
2
Weak: only partly works, with substantial simulated functionality.
1
Does not meet the criterion.
How the final score works
Final score = Σ (score × weight) ÷ 5
Maximum: 100 points. For example, Technical execution 4, Usefulness 5, Originality 3 and Demo quality 4:
(4 × 40 + 5 × 25 + 3 × 20 + 4 × 15) ÷ 5 = 81
What counts as a working build?
AI agent
Counts
Decides its next steps without being prompted at every step.
Calls real tools, APIs or databases and uses their results.
Has a loop: attempt, fail, correct and retry.
Completes a task end to end with a real output.
Handles errors when a tool fails.
Does not count
A question-and-answer chatbot that stops after an LLM response.
A chat interface wrapped around a single LLM call.
A linear flow that only passes text along.
Advice that leaves the user to carry out the task themselves.
A build that crashes as soon as an API fails.
If you remove the LLM, is there still a system left? A chat box alone is a chatbot, not an agent.
Workflow / automation
Counts
A real trigger leads to real output: an email sent, a database updated or a document created.
Runs again in front of the judges with consistent results.
Processes data arriving during the demo.
Handles failure with retries, fallbacks or notifications.
Connects to a real system such as a sheet, database, WhatsApp, email or partner API.
Does not count
A flowchart or process diagram.
Manually running each node while explaining it.
Hardcoded dummy data.
A build that works only on one perfect scenario.
A flow that stops at the editor canvas.
The judges provide a new input, the workflow runs, and something real changes at the other end.
IoT
Counts
A physical device powered on in the room and transmitting real-time data.
Sensors reading real conditions; judges can interact with the sensors.
An action back to the device through an actuator, relay, alarm or motor.
Sensible handling of a lost connection.
An original sensor combination, placement or use, with a clear reason a physical device is necessary.
Does not count
Sensor readings simulated in code.
A video of hardware running somewhere else.
A dashboard showing static numbers.
A build that needs perfect Wi-Fi to work.
A standard board and tutorial with no original contribution.
Why must this be physical? If an ordinary phone app solves the problem, the innovation score is lower even when the circuit is tidy.
Product (B2C / B2B)
Counts
A URL the judges can open on their own phone or laptop.
Real accounts and data, beyond one pre-filled screen.
A complete core journey from start to finish.
A build that handles unexpected input.
Does not count
Figma prototypes or click-through mockups.
Screenshots inside slides.
A polished landing page with buttons that do not work.
A build that only works with rehearsed examples.
The judges open the URL themselves, provide new input and complete the core journey.
Bring your skills, not a finished codebase.
The core of what you demo must be built during the 24 hours. Boilerplate, libraries and your own reusable utilities are allowed; declare them in your submission summary. Demo videos are limited to 3 minutes and do not replace the live run.
Be upfront about unfinished parts. Judges score what runs. If hardware fails, tell the judges immediately and debug what you can in the room; switching to a recording does not satisfy the live-run requirement.