.png)
☰☰Three projects turned hyperlocal temperature intelligence into working tools for extreme heat. Here is who won, and what they built.
Extreme heat is one of the fastest growing risks to cities, and most tools still treat temperature as a single number for a whole city. FortyGuard models it at the scale people actually feel it, street by street and hour by hour. FortyGuard Hackathon'26, Building the World's Temperature AI, asked developers to do something real with that.
Over two weeks in August the challenge ran fully virtual and fully global, built on the FortyGuard Temperature API® and sponsored by NVIDIA. More than 3,800 builders from over 50 countries and six continents took part, forming hundreds of teams and shipping 556 working projects. Participants rated the experience 4.5 out of 5.
Every entry was scored on impact and relevance, technical execution, innovation and communication. After screening and a full round of judging by a panel drawn from climate, technology and venture, three projects rose to the top.
Each winning team received a cash prize and an NVIDIA Jetson AI Developer Kit.
Team Ministry of Temperature: Abdul Moiz Zahid Siddiqui, Azaz Ur Rehman Nasir and Muhammad Faizan Raza (NYU)
The problem. Cities decide where to spend cooling budgets with very little to go on. Heat is usually reported as one figure for a whole district, so the choice of which building to shade or retrofit, and which floors and facades, comes down to guesswork.
How it works. The Urban Canopy is a facade scale physics model of Midtown Manhattan. Instead of one value per block, it models 29,415 individual facade panels across 5,329 buildings for every hour of the year, anchored to measured FortyGuard air temperature. For any building it produces a costed intervention brief: which measure to use, on which floors and facades, what it saves and what it costs. At city scale it ranks buildings by heat exposure and by the people behind those facades, then allocates a capital budget. In the team's model, $57 million selects 91 buildings and 292 measures and cuts exposure by 1.88 million person hours above 35°C. An in app analyst, built on the Claude Agent SDK, answers questions by rerunning the physics and showing its reasoning.
Why it stood out. It took temperature intelligence all the way to a decision a city could actually act on, at the scale of a real skyline, and paired the model with an interface that explains itself.
Built by Bismah Javed
The problem. Data centres often run mechanical chillers when the outside air is already cold enough to cool for free, because the plant needs advance notice to switch and cannot tell what the intake air will do. Operators hold a fixed safety buffer and overcool as a matter of policy.
How it works. Agentic-Arbiter is a free cooling scheduler. It turns FortyGuard's two metre air temperature forecast, the height a ground mounted condenser actually breathes, into an hour by hour schedule that tells a chiller plant when it can safely cool on outside air. Each hour carries a safety margin measured from the agent's own past errors rather than a fixed figure, and where the site geometry defeats the physics it declines the hour instead of guessing. It models the exhaust plume on each building's real footprint using NVIDIA Warp, and runs across 238 real data centre campuses in 36 states.
Why it stood out. It was the only entry to publish its own shortfall as openly as its results, reporting live coverage against target and explaining the gap. That honesty about a model's limits was exactly the kind of rigor the judges called out.
Built by Darin Levesque
The problem. Frost can wipe out a vineyard harvest in a single night, and growers place wind machines, the fans that pull warmer air down over the vines, largely by intuition.
How it works. FrostLine turns that placement into a data driven plan. It pulls multi year cold hour history from the Temperature API, scores risk against the grapevine growth stage most vulnerable to frost, and recommends where machines should go, then estimates acres protected and crop dollar savings against install cost, with every assumption shown. The team deliberately pointed a model built around heat at cold as a stress test.
Why it stood out. When the readings ran warm against a real April 2026 Virginia freeze, the team did not hide it. They built an independent cross check into the tool and reengineered the scoring to rank risk relatively, so it improves automatically as the underlying model does. Turning a hard result into a better product is what set it apart.
Hackathon'26 was built by its community. Thank you to every builder who gave up two weeks to turn temperature into something useful, to the mentors and judges from across NVIDIA, Google Cloud, Autodesk and beyond who gave their time, and to NVIDIA for sponsoring the event.
The recordings from the mentor and judge sessions are on our YouTube channel, and they will be available here on the hackathon page later this week.
This is only the start of what gets built on temperature intelligence. Congratulations again to The Urban Canopy, Agentic-Arbiter and FrostLine.
.png)
.png)
.png)
.png)
.png)