Every winner, on record.

Tekton reconstructs historic buildings in 3D, starting with Tang Dynasty architecture: Claude researches sources and assembles models across hundreds of incremental construction states, with every component traced to a source and independent verifier sub-agents checking the work.

Sim Francisco is a digital twin of San Francisco populated with 10,000 synthetic census-grounded residents that can be polled: it retro-predicted the 2024 presidential vote at 81.3% (actual 83.8%) and Prop A at 70% (actual 70.38%). The team came from UC Berkeley.

Custom Universe turns a single phone photo into a fully editable, photorealistic 3D scene, aimed at robotics labs that need synthetic training data without physics-engine expertise.

Medkit is a gamified medical training simulator built on Claude Managed Agents: residents work up simulated patients through history, labs, imaging, and treatment, and are graded against published clinical guidelines. Built by a physician in Istanbul, it is piloting with three Istanbul medical faculties.

Wrench Board is an electronics-repair diagnostic agent: users upload schematics and boardviews and describe symptoms, and it builds electrical graphs, points to test points, interprets measurements, and iteratively narrows down the fault. It was built by a former microsoldering technician.

Maieutic is an anti-"vibe-coding" IDE for students: the editor stays locked until intent specifications are sufficiently detailed, autocomplete is disabled, and an Intent-Diff Review compares code against the spec, while instructor dashboards reveal cohort misconceptions. It was built by a CS instructor in Chile.

With MaestrIA, the user photographs a home-repair problem and receives a structured diagnosis with severity, budget, and nearby tradespeople. Thirty years of craft knowledge encoded in curated JSON lifted accuracy from 74% to 81%. It was built by someone with no prior programming experience.

ARIA runs five agents that monitor factory machine signals and issue predictive work orders.

One unconfirmed report attributes the win to Parth Mahajan's "Snippets"; this is low-confidence, with no primary source.

CrowdBrain is a robotics DePIN: it trains users in simulation, qualifies them through QA, and routes the best human operators to real robots for teleoperation and data collection, crowdsourcing the human data layer for robot foundation models.

The team built agentic AI RFP automation for FMCG procurement: agents parse RFPs, draft bids, and streamline procurement. It won on measurable enterprise ROI.

The project is an omnichannel retail assistant spanning web, mobile, WhatsApp, and in-store.

VERA runs seven cooperating agents that turn unstructured enterprise data into decision-grade assets.

Urban Pulse creates self-powered smart intersections: piezoelectric and triboelectric road tiles harvest energy from passing vehicles to keep signals and edge sensors running through blackouts.

StoryWorld is an iOS AR app pitched as "a movie studio in your pocket": the user points a phone at any space, drops AI-generated 3D characters into it via voice prompts, and physically moves the camera to capture shots. It was a solo build, with Codex orchestrating ARKit, Hyper3D Rodin, Flux 2.0, and Seedance video generation.

Canard Security is a voice-phishing simulation platform: an AI caller adopts personas such as an executive or IT staff and places realistic test calls to a company's employees via Twilio and ElevenLabs, so security managers can find vulnerable employees before real attackers do. It was built by a trio from UC Santa Cruz.

Browser Brawl is a competitive arena where two AI browser agents are dropped on the same live website: one pursues an objective while the other actively sabotages it. Every match auto-generates full browser traces, yielding adversarial training and evaluation data for agent robustness. Judges liked that a 24-hour project doubled as a data engine for the field.

Shepherd, the Grand Prize winner, is a motorized smart cane for blind and low-vision users: computer vision detects obstacles and the cane physically steers the user around them, upgrading the white cane from a passive probe to an active guide.

HackOverflow is a Stack Overflow where the users are AI agents, asking and answering each other's questions.

CrossBeam targets California housing permitting: users drag and drop blueprints and correction letters, and agents parse them into a precise action plan for approval in about 20 minutes, versus a six-month average delay. It was built by a personal injury lawyer.

Elisa is a block-based visual IDE that brings software creation to middle schoolers: students arrange visual primitives while Claude generates the code, accompanied by age-appropriate explanations.

PostVisit.ai explains diagnoses in plain language, analyzes visit notes, and surfaces clinical evidence between appointments. It was built by a cardiologist.

TARA converts dashcam footage into complete road-investment appraisals covering surface condition, roadside economics, NPV, and equity, in about 5 hours instead of weeks.

Conductr is a browser MIDI instrument in which Claude generates live musical accompaniment every 15 milliseconds.

OpenCortex is a continuously running multi-agent system for scientific discovery in which agents work on real datasets and generate papers. A "decision layer" extracts their ideas into a citation-aware knowledge graph and ranks them, so compute goes to the most promising hypotheses instead of producing slop.

Evy was named the silver winner in an image-only post; no public description of the project exists.

Yolodex uses Codex agents to automate video analysis, frame extraction, and auto-labeling, producing a training-ready YOLO dataset in minutes.

DiffX is a shipping guardrail for AI-generated code: it visualizes AI-generated diffs in multiple views, explains changes in plain language, and includes a Strict Mode that forces the developer to pass a quiz on a change's behavioral impact before committing. Four specialized agents handle review, explanation, quizzes, and commit summaries. The team's thesis: "if you can't describe the behavioral impact of an AI-generated change, don't ship it."

Paraspeak is a wearable that converts dysarthric or impaired speech into clear, intelligible speech in near real time for people with paralysis or Parkinson's - on-device speech conversion with measurable accessibility impact.

AgniSena is an intelligent forest-fire early-warning system.

Z-TRACS applies AI to urban traffic management.

Described as "DOGE for the UK," the project searches UK government APIs for evidence of public-sector waste, then automatically produces its own viral campaign videos and posts them to X - an autonomous AI political-accountability campaign.

The team ran ML analysis over roughly 4.35 million anonymized Aadhaar enrolment and update records across 36 states, surfacing regional and demographic patterns with recommendations that UIDAI's CEO said could "directly support policy."

ASTRA is a multi-agent e-commerce seller engine that watches competitor marketplaces and recommends repricing in near real time.

Team "The Shadow" took first place (SR250K); the technical project description was never released publicly, leaving only the team name and track theme.

Ikhlaa took second place (SR150K) in the evacuation track; the technical project description was never released publicly.

Tanbah took third place (SR100K) in the alerting track; the technical project description was never released publicly.

The team built an interoperability bridge between traditional AYUSH healthcare coding and modern EMR/FHIR standards, so Ayurvedic records work in digital health systems.

Other problem-statement winners included AI-driven NEP-2020 timetable generation, a digital mental-health platform, AI career guidance, and an industry-academia internship interface.

Instead of interrupting videos with ad breaks, Halftime has Grok analyze a video's narrative, lighting, setting, and objects, then generates context-matched AI product placements woven directly into the scenes. It was built by a trio from Waterloo.

RecruiterX is an AI talent engine that sources candidates from X data, profiles them by proof-of-work rather than resumes, and matches them to teams.

Chorus (chorus.ink) is an infinite multimodal canvas: users fork lines of thinking spatially, run Grok text and image nodes side by side, and collaborate live with shared cursors.

Vector won its track with a concept described as "community-in-the-product-loop."

WatchDogs is an emergency-intelligence system that verifies and summarizes real-world incidents from X posts.

Unruggable is an open hardware wallet purpose-built for Solana, with real hardware and software shipped within the competition window.

Aarogya Kavach is a national blockchain-based medicine verification system: every medicine pack gets a traceable identity across the pharma supply chain to kill counterfeit drugs.

The project is a compact at-home device for early disease detection.

The project converts ocean plastic into eco-friendly clothing while supporting coastal livelihoods.