OpenAI Reveals 8 Build Week Award Winners

Total prize money came to $100,000. Each category awarded $15,000 for first place and $10,000 for second, with first-place teams also receiving DevDay passes, a meeting with the Codex team, and a year of ChatGPT Pro. Judging was based on four criteria: technical execution, design and usability, impact, and idea quality.

The four categories were Apps for Your Life, Work & Productivity, Developer Tools, and Education. Each category revealed one first-place winner, one second-place winner, and three finalists.

Apps for Your Life — giving voice to what couldn't be said

First place, Second Voice, was built by Ravitez Dondeti, who had a relative with a speech disability. As a child he communicated with her through gestures, and only after she passed away did he realize how much he'd never gotten to ask. He built the app alone during the eight days of Build Week.

For people with dysarthria or motor limitations, the app overlays a personal phrasebook and real-time context onto fragmented speech, surfacing just a handful of likely sentences. Only after the user selects or edits one does the app speak it aloud. The person always keeps final control over what actually gets said.

Dondeti expected sentence reconstruction to be the hardest part. What actually held him up were the smallest decisions: how many options to show, how fast to surface them, and how many taps a confirmation should require.

"The smallest UX decisions carried the most weight. An app can work perfectly on the technical side and still be useless if people can't actually use it."

Second place, AirBridge for Windows, grew out of a simple frustration: there was no way to use home AirPlay speakers from a Windows PC. Adam Tarantino had tried solving this once before with a pre-2024 OpenAI model but never got it to a working product. This time he went back at it.

The app streams Windows audio to HomePods, Apple TVs, and Macs without virtual audio cables or temporary files. It handles simultaneous playback across multiple speakers, room-by-room latency correction, and even a browser extension that delays video instead of audio to keep lip sync intact. GPT-5.6 handles voice control of the system, a local policy layer defines what actions are allowed, and results get verified against real hardware.

"I've been building software for over a decade, but this was my first hackathon. You can't win if you don't show up."

Three finalists rounded out the category. Bander lets families review what an AI assistant plans to do before it touches an account. O2 by Agent9 brings air quality, wildfire data, alerts, and forecasts into a single view. SayAhead helps deaf and hard-of-hearing users read, steer, and control phone calls.

Work & Productivity — when three vets become one

This year, Pauli VetCare went from three veterinarians to one. Client calls kept coming in, but there was no longer enough staff to see every case.

First place, veTriage, was built by Erin Downes, the 61-year-old veterinarian who runs that clinic — with no coding background. The tool helps front-desk staff ask the right medical-history questions, catch red-flag symptoms, and route cases to the appropriate review path, while making sure non-clinical staff never make the actual medical call. The core insight: a packed schedule doesn't make a patient less urgent, so clinical need and clinic capacity need to be handled separately.

Codex turned Downes's clinical judgment into a working workflow, and GPT-5.6 supports the intake conversation without making medical decisions. Downes's clinic is now piloting the tool in real operation.

"At 61, what I want people to know is that it's not too late to become a builder if you have deep domain knowledge."

Second place, Pulse, is a research prototype built by Mohamed Mostafa Abu Taleb, a cardiologist in Cairo, in the gaps between hospital shifts. During cardiac arrest resuscitation, the team leader has to simultaneously track rhythm and defibrillation, medication timing, CPR cycles, and interruptions to chest compressions — all while people shout information from every direction.

Pulse listens to the room and maintains clinical state, accounting for speech that mixes Egyptian Arabic and English. GPT-5.6 interprets the chaotic speech, auditable deterministic code tracks the resuscitation procedure, and the system asks the team for clarification whenever the evidence is unclear.

"No team, no lab, no funding — just built in Cairo, between hospital shifts."

Three finalists: LabSpace AI turns a lab and everything in it into a searchable spatial map. OpenCounsel transforms broken legal drafts into properly sourced, filing-ready documents. Tomok stitches together schedules, specs, and reports so infrastructure teams can trace the source of delays.

Developer Tools — sketching sound before building the room

Spatial audio is usually only heard after an entire game-engine scene has already been built, meaning there's no way to explore "how should this space sound" early in the design process.

First place, Echo Canvas, flips that order. Built by Kevin Yang, the tool lets designers sketch a space directly in the browser — the same way a web designer blocks out a layout with gray boxes before picking colors — place a sound source and listener, and hear the result instantly as they open a door or change a wall's material. That's why he calls it an "acoustic wireframe."

Geometry and acoustic calculations run through a deterministic system, and audio renders locally in the browser. GPT-5.6's role is limited to composing and describing scenes within a constrained schema.

"Firing more rays doesn't make a better product."

Second place, Sentinel, is a security scanner for MCP servers, which give agents access to files, APIs, databases, and shell commands. Malik Bashaar Javaid kept running into MCP servers shared as "quickstart templates" that left raw shell calls, hardcoded credentials, and weak authorization boundaries untouched.

Sentinel runs deterministic static analysis alongside a tightly constrained GPT-5.6 review that operates only within real source context, plus Docker-isolated probes. The model can support or challenge a finding, but it can't fabricate an execution probe or cite code that doesn't exist. Results map to the OWASP Agentic Top 10 and can feed into GitHub code scanning.

"Constraining the model was harder than writing the prompts." It was his first hackathon, right after finishing a computer science degree.

Three finalists: Emberframe Studio is a spatial canvas for building with AI agents. GenUI converts model-written JSON into verified native SwiftUI. Vibe Signal lets users monitor and direct Codex from an iPhone or Apple Watch.

Education — the wall that had blocked a 25-year developer

First place, Mechanica, started in a museum. "I don't know how many times I've bumped my head against the glass — I get too absorbed in what's behind it," says Shan Wei. Looking wasn't enough; he wanted to touch it, take it apart.

The team brought four types of ancient Chinese machines, known only from a handful of surviving records, back to life as working reproductions — from an astronomical clock tower to a programmable loom that predates modern computing by centuries. Visitors can operate and disassemble the machines, tracing every dimension back to a classical text, a surviving artifact, or explicitly labeled scholarly inference. Where scholars disagree, the project shows competing reconstructions side by side instead of picking one answer.

Weiying Zhu, Yukun Li, and Shan Wei all had long software careers but had never touched 3D modeling, physics simulation, or animation before. Codex carried them across that unfamiliar territory, and GPT-5.6 became an AI docent that cites the museum's sourcing or declines to answer when no evidence exists.

"I've been building software for 25 years, but 3D and animation were completely outside my world — and that barrier just disappeared."

Second place, Dấu, was built by Robert Huynh, who once walked out of a previous hackathon after half a day, convinced his idea wasn't "technical enough." This time, he showed up solo in Hanoi.

In Vietnamese, the same syllable can carry six different meanings depending on pitch and voice quality. Dấu lets learners record a word, plots their pitch curve alongside a verified native-speaker reference, tells them what meaning their pronunciation actually conveyed, and gives specific physical corrections. Deterministic signal processing handles tone detection, while GPT-5.6 only coaches. When the evidence is unclear, it asks the learner to try again rather than confidently guessing wrong.

"I went from wondering if I was technical enough to even join a hackathon, to building something solo that became a finalist."

Three finalists: Canopy lets learners immediately apply what AI just taught them in a coding sandbox. Encore! turns a child's missed test question into a custom picture-book lesson. ResearchOS lets users trace every research claim back to its source.

Winners at a glance

Category	First place	Second place
Apps for Your Life	Second Voice — turning fragmented speech into sentences	AirBridge — streaming Windows audio to AirPlay
Work & Productivity	veTriage — phone triage for veterinary clinics	Pulse — tracking cardiac arrest resuscitation
Developer Tools	Echo Canvas — acoustic wireframing	Sentinel — security scanner for MCP servers
Education	Mechanica — reconstructing ancient Chinese machines	Dấu — a Vietnamese tone coach

Editor's take

Line up all eight winners and two patterns jump out.

First, the winners weren't the best coders — they were the people who understood the problem. Two of the four first-place teams were a veterinarian and a cardiologist. The other two crossed into unfamiliar territory of their own — 3D modeling and spatial acoustics — using Codex to get there. Hackathons used to be a contest of raw development speed. Here, the axis tilted decisively toward domain knowledge.

Second — and this matters more for practitioners — most of the winning projects kept AI on a short leash.

Pulse handed resuscitation-procedure tracking to deterministic code and gave the model only speech interpretation. Echo Canvas built its own geometry and real-time audio engine, using the model purely as an authoring and explanation layer. Dấu used signal processing for tone detection and limited the model to coaching. The Sentinel builder said outright that most of his work wasn't prompting — it was building constraints.

Anyone who's worked extensively with coding agents knows exactly where this usually breaks down. The wider the scope you hand the model, the flashier the first demo looks — and the shakier the results get on the second run. In practice, teams end up spending most of their time narrowing what the model is allowed to touch, using schemas and rules.

That's exactly what these winning projects are saying, in unison: let human-written code own judgment, computation, and procedure, and confine the model to the parts that involve understanding and explaining what a person said. And when the evidence is ambiguous, have it ask again instead of making something up.

The fact that eight winners, picked out of 8,000 submissions, all converged on the same approach isn't a matter of judges' taste. It looks a lot closer to an actual answer for how this technology should be used right now. The model earns its value as an adapter, not an engine.

The same logic applies directly to teams here. Organizations that know their ground truth — hospitals, warehouses, legal departments, manufacturing floors — are in the strongest position right now. What they need isn't more developers on staff; it's putting Codex in the hands of someone with 20 years of frontline experience and giving them two clear weeks. That's exactly how veTriage was built, and it's now running in a real clinic.