Winners Gallery

An eight-agent knowledge engine that stops institutional knowledge from walking out the door — router, knowledge graph and dashboard, with 43 tests. A judge called it 'a product, not a hackathon project'.

Works inside code review: spots the security bug, writes the fix, and opens the merge request — no human needed.

Explains every decision it makes with AI, deployed to Google Cloud and production-ready.

Maps code dependencies and shows how systems evolve — see a change's full impact before you make it.

Three agents — Detector, Writer, Reviewer — fix documentation; confident fixes become merge requests, uncertain ones become issues for humans.

Five agents that run database migrations against a safe copy of production — real PostgreSQL, real data, real Google Cloud deployment.

Verifies AI-generated security reports, closing the trust gap between AI findings and developer action.

Scans CI/CD pipelines for environmental impact and produces carbon footprint reports.

Describe what you want to ship in natural language; Wayfinder gathers freight rates, tariffs, regulations and logistics risks in one place.

Makes AI-generated intelligence cryptographically verifiable — every source fingerprinted with SHA-256 and anchored on Solana.

A 24-agent parallel probe that detects and classifies location-, hardware- and cookie-based price discrimination in real time.

AI-visibility intelligence for marketing teams: visibility score, citation gap analysis and grounded recommendations.

Listens to 125 regulators across 67 jurisdictions — sites, PDFs, hearings — and maps every change to SOC 2 / ISO / NIST controls.

Demand-side early warning for commercial real-estate rents — detects market turns 5–7 months ahead, 79.3% hit rate across 17 metros.

AI exam-prep companion for K9–K12 students across the EU, launching in Estonia with national-curriculum diagnostics.

A multi-agent system connecting female hormonal biology to executive decision-making — bio-intelligence for women who lead.

Autonomous trading agent that ranks signals, backtests, exposes risk controls and keeps a read-only audit trail.

Detects mental-health spirals before they deepen: each memory is stored in Qdrant as six named vectors (semantic, emotion, prosody and more), a custom transformer reads 14-day emotional sequences, and when trouble looms it retrieves the user's own past recovery as proof they have climbed out before.

Simulates how hundreds of listeners react to a DJ set in real time: every persona lives in Qdrant as a 532-dimension hybrid vector, scored every 15 seconds with drifting preferences, while the crowd's convergence or polarization is projected onto a 2D map.

A dark-fantasy RPG where every player decision becomes a vector 'soul history' — NPCs quote your exact past deeds in dialogue, and story gates unlock based on accumulated episodic memory, all powered by Qdrant Cloud.

An autonomous vendor-contract compliance system: a Copilot Studio orchestrator and four specialist agents score inbound contract PDFs against a 15-rule Dataverse rulebook with red/amber/green verdicts, going from email to report in under two minutes — the highest-scoring entry of the entire hackathon.

A Copilot Studio agent that turns performance-review feedback into a concrete development roadmap — pulling skills from SharePoint, surfacing Microsoft Learn resources, matching internal mentors, and routing manager approval through Power Automate.

An AI scrum assistant: forward an email with meeting transcripts and it extracts sprint goals, epics and stories, enriches them via the Microsoft Learn MCP server, and files issues in Jira Cloud through a Jira MCP server — reportedly cutting sprint prep time by about 70%.

An autonomous IT service desk for small businesses: a SharePoint portal feeds a Copilot Studio triage agent via Power Automate, managers approve with one tap in Teams, and actions execute through least-privilege Microsoft Graph calls — with the whole stack built and maintained by AI agents under a single person's supervision.

A privacy-first Android app that analyzes toddler walking videos entirely on-device: it checks capture quality, runs 33-landmark pose estimation locally, applies deterministic gait analysis, and has a local AI assistant explain the results to parents (a research prototype, not a diagnostic tool).

An affordable wearable that turns a phone camera into directional buzzes on a belt, steering blind and low-vision users around obstacles — fully on-device, no internet required.

An offline survival copilot for when GPS and networks fail: on-device first-aid triage, a solar/star compass that finds true north, and medic-to-casualty translation — all running locally.

A FinOps platform whose agents hunt down wasted Kubernetes spend on their own — spotting it, diagnosing the cause and applying the fix. Built on kagent, kgateway and kmcp, it turns cost optimization into something that runs continuously instead of quarterly.

A Kubernetes-native platform that federates several MCP servers behind a single endpoint, letting IDEs and AI clients query MIMIC-IV clinical data securely through agents, tools and service discovery.

A Kubernetes-native, multi-cluster inventory controller for enterprise AI infrastructure, extending the agentregistry idea so teams can actually see what agent infrastructure they are running at scale.

An agent that gives kagent eyes: it grabs a webcam photo and transforms it with Nano Banana and Veo3. It also ships an American Sign Language mode, so accessibility was part of the design rather than an afterthought.

A bridge that wires kagent to the Microsoft Agent Framework over A2A, so agents built on the .NET side can be brought into the kagent ecosystem as-is.

Nine substantial pull requests that pushed agentgateway from a model proxy into an identity-aware control plane for agents — identity at the edge, OpenTelemetry visibility into agent execution and support for multiple model providers.

Kubernetes-native governance for MCP infrastructure: it watches agentgateway and kagent resources, scores security posture per MCP server, and explains the risks using Gemini or Ollama.

A full-stack e-commerce site built to demonstrate MCP authentication and authorization. agentgateway, Keycloak JWTs and CEL rules are enforced at the protocol layer, so agents cannot route around the security boundary.

Re-ranks the patient queue by clinical urgency inside OpenMRS. Rather than leaving triage to a generative model, it uses the deterministic NEWS2 score, so the behaviour is predictable — judges and community voters both put it first for how deployable it looked.

An offline-capable stack that transcribes and translates a consultation on the spot. Audio never has to leave the site, and the workflow assumes a clinician reviews the output before it is used.

Supports record-keeping and situational awareness at disaster medical sites, working from the conversations and wound images that occur there. The team identified where LFM fell short, prepared data and fine-tuned to close the gap — unusually thorough execution for a hackathon window.

Detects the signs of a phone scam in real time and warns the person while the call is still happening. It targets a problem specific to Japan, and the judges credited the choice of where to put the LFM and the care taken over the user experience.

Drop in a long broadcast video and it works out the edit plan, picks the scenes and writes the captions to produce a social-ready short. Afterwards you can still adjust it by chat — 'make the captions yellow' — and it took the top spot against 28 teams from Japanese broadcasters.

A 3D anatomy atlas driven by plain language. Ask about cerebral infarction and the AI picks the relevant organs and vessels, renders them in 3D and adds an explanation. Interpretation, identification and drawing are separate stages, holding down wrong displays, and ambiguous input comes back as candidates to confirm.

An agent that builds a Minecraft town from a vague description. It draws a concept image, works out zoning and building plans, then converts the image in two steps, putting the structure into language before turning it into coordinates, which keeps the result coherent. The frame goes up in bulk; a bot adds decoration on site.

An app that lays the laws behind everyday scenes over the scenes themselves in AR. Pick a difficulty and a field, point a phone at a fountain or a bridge, and an AI reads the image and draws trajectories and force vectors on top of what the camera sees. If curiosity goes further, a link leads to a fuller written explanation.

An app that suggests what you can build with the bricks you own. It recognises parts from photos to build an inventory, and once you pick a theme and difficulty the AI generates a 3D build sequence. A physics engine checks it for overlaps and floating pieces, reports the faults and makes the AI redo it, up to three times.

An experimental system that clears intersections by market instead of traffic light. Each vehicle's AI agent negotiates for right of way, weighing urgency and its funds, and settles with a micropayment. Small instant payments run without handing over keys, sketching traffic as it might look once driving is autonomous.

A device for talking to a character reincarnated as a plant. Sensors watch soil moisture, temperature, humidity and light, and pass the readings to the AI as feelings — thirst, cold — so it speaks up with a face and a voice. It starts talking when a motion sensor spots someone, so a child gets a cue to care, no numbers needed.

A platform linking museums and visitors. From a painting's image, AI writes commentary at three difficulty levels in several languages, delivered as an audio guide. Preferences left on each work are matched against embeddings to suggest what to see next and where it hangs, while museums see a per-work view of how pieces land.

An analysis platform for baseball teams short of both coaching and money. Pose estimation from video turns a player's movement into numbers, replacing instruction based on feel with feedback in figures and words. It also fields position-specific questions and suggests training menus, on a serverless setup that keeps costs low.

An AI partner for finding things around the house. It watches the room at intervals, keeping a log of footage, so asked where the keys are it reports not just a place but a state, such as hidden under a magazine. If they are absent, it reasons from calendar entries to suggest candidates, and turns its camera with a servo.

In the US the help that follows release from prison is real but fragmented: housing, identification, legal aid and employment programs each sit with a different organisation. Alyosha puts returning citizens, providers and caseworkers in one place and walks them through those pieces as a single route back into ordinary life.

A benefits navigator that works in the browser and speaks three languages: English, Spanish and French. Users type or talk through their situation, NLP pulls out what they need with confidence scores attached, and the tool then walks them through what they may qualify for and keeps checking back until help has actually arrived.

A decision-support tool for the people who have to answer the opioid crisis on the ground: county health directors, lawmakers and advocacy groups. It shows how large the problem has grown, guides the user through shaping a response, then puts a price on that response alongside the savings it could produce.

Built so that municipal offices in Peru can see a disaster coming while there is still time. Seven AI agents cooperate, scanning in real time for earthquakes, floods, fires, unusual weather and news coverage, and pushing an automatic alert to local officials before conditions escalate.

A multi-agent release gate that asks a different question: what could this pull request break in the real world? Rather than stacking up pass/fail checks, it evaluates the change in a production-like context and returns a deployment confidence score from 0 to 100, together with a risk brief written in plain English.

A multi-agent system for e-commerce sellers whose catalogues are too large to watch by hand. It keeps an eye on competing marketplaces, recommends repricing moves and can apply them automatically. What it targets is the margin that quietly bleeds away while pricing and risk decisions lag behind what the market is already doing.

Puts live road conditions in front of the people running fleets and logistics operations, so routing and delivery decisions follow the current traffic picture rather than guesswork. Built on the Microsoft AI Platform, it combines that traffic view with fleet visibility and optimisation hints to cut supply chain delays.

Shows a team what is really running across their Azure estate and what deserves attention first. It keeps discovering resources, scores them against architectural and operational best practice, and then proposes concrete next steps aimed at reliability and cost.

An AI code review extension that lives inside the Azure DevOps pull request flow. It reads each change automatically and flags security risks along with policy and compliance gaps, and it does so early, while the branch is still open, so developers deal with the findings before the merge rather than after it.

An AI-powered SDR system: automated lead generation, research, proposal writing and outreach. Grand prize among 10,400 participants from 62 countries.

Multi-agent AI transforming energy customer management through ADK orchestration.

Democratizes education in Brazil: autonomous agents grade essays, build personalized study plans and generate mock exams.

An AI team that continuously audits, forecasts and optimizes cloud infrastructure for sustainability.

Next-gen personalized education: interactive lessons with visuals, quizzes and smart AI support.

Converts natural language into validated Feynman diagrams using real physical laws — bridging theory, automation and symbolic reasoning.

Multi-agent financial analysis platform that evaluates trading hypotheses, built on ADK and Vertex AI.

A visual AI agent builder: describe agents in plain English, design visually, test instantly.

A quiz game weaving the theme 'Light' through varied scenes, by team Kamomi — won on overall completeness in live demo judging.

Built by Kamie Shinju in a 45-minute sprint from the concept of 'excuse' — won the division on combined team and solo results for reliably working output.

AI expert-matching for organizations: builds a knowledge graph from Jira, GitHub, internal docs and resumes — you can even query a person's digital twin before reaching out.

AI transaction analyzer that detects fraud, alerts users by email and takes reports via an AI voice assistant, using Llama API multimodality to cross-check client news.

A CCTV control-room AI operator with zero fine-tuning: define surveillance events in plain language and Llama 4's multimodal vision detects and reports them.

Processes 400-page geology reports with Llama 4 Maverick's long context and generates 3D subsurface models — built by a working geologist.

Turns e-commerce discovery into an explorable 3D world — categories become mountains, result counts become terrain height, and an LLM guide narrates the tour.

A trajectory memory for robots: banks sensor-action sequences as vectors labeled by safety, then retrieves the safest neighbor to suggest the next move.

Game NPCs that remember places and events via spatio-temporal memory — CLIP image memories plus text memories, running locally on ~3GB VRAM at 1-2s latency.

Route-aware event discovery: enter a route and time window, and hybrid semantic search plus geospatial-temporal filters surface events along the way.

An AI choreographer that generates personalized Bachata sequences by matching music features to labeled dance clips.

Transforms cosmetics retail: analyzes customer data across store, EC and counseling touchpoints to generate personalized recommendations that close the gap between veteran and novice staff. Won among 37 teams.

A team of AI agents dividing strategy, copywriting and analysis inside Slack to run the marketing PDCA loop — 'a parallel AI workforce'.

A regional-revitalization agent that bridges municipalities, companies and prospective migrants — matching regions, jobs and programs through to post-move settlement.

The Robohon robot chats with elderly users via Agentforce, auto-writes daily reports shared to family on Slack, checks medication and flags health anomalies.

Analyzes your grocery cart and recommends recipes, built on Gemini, GKE Autopilot, ADK and A2A. Grand prize among 4,773 participants from 133 countries; the winner spoke at KubeCon.

Reinvents credit pre-approval with a multi-agent pipeline that analyzes spending patterns and balances bank profitability with customer value.

A plug-and-play AI microservice for online stores: image mixing lets shoppers see themselves wearing the product, plus smart descriptions for accessibility.

Intelligent retail automation with personalized chat, virtual try-ons and instant ad generation, orchestrated on Gemini and GKE.

Conversational shopping through six specialized AI agents that understand style and learn preferences, deployed as a multi-agent architecture on GKE.

An automated pipeline that ingests raw technical demo videos from YouTube, distills their content into a structured knowledge base, and makes it searchable. Turning scattered video into queryable data won it first place in the inaugural Free Edition hackathon spanning 16 countries.

A space-weather analysis system for power grid operators: it predicts grid failures triggered by solar flares and serves up 7-day forecasts, risk thresholds and recommended actions, applying data engineering and ML to critical-infrastructure resilience.

A recipe recommender that structures raw recipe data with NLP and clusters dishes by theme and flavor profile, so users can ask in natural language and get personalized suggestions. Took third place as a practical showcase of text-based workflows.

A real-time ambulance routing engine on Microsoft Fabric: dispatch events landing in a mirrored Azure SQL database trigger a notebook via Activator, which combines the Google Routes API with an ML model predicting the speed advantage of siren privilege, surfaced to control rooms on a Power BI map. Built by an Italian developer.

An AI system that makes city waste collection smarter: it monitors bin fill levels, predicts when they will overflow, and dynamically reshuffles collection routes. Built by a Kenyan team with Fabric Data Agents, Copilot Studio and Azure OpenAI.

A composition aid for school music classes. A student describes in their own words the song they want to make, and the AI proposes chord progressions and melodic techniques that could fit. It stops short of writing the piece itself: students try options and choose, bringing individual creativity into uniform-feeling lessons.

A tool for primary and junior high teachers who lose hours each week writing class newsletters. Teachers just talk to it: the AI asks back for missing details, assembles the draft, and outputs HTML or PDF laid out to survive A4 printing. Because it keeps the paper habit intact, even teachers uneasy with IT can adopt it.

A declarative travel planner that starts from a mood rather than a destination list, say wanting to kayak on a quiet lake, and builds an itinerary from it. It reads across popular spots, hour-by-hour crowd data and reviews, then orders stops to dodge the busy hours. Change your mind mid-chat and it recomputes on the spot.

A shopping app for when you cannot tell if furniture will fit your room. AI reads dimensions and existing pieces from photos and builds a 3D digital twin. Products drop into that twin for checking, and because the AI knows what is on screen and where things sit, you can ask for advice, or a search, without explaining.

A service that hands quotation work, long the preserve of veteran staff in woodworking shops, over to AI agents. The job is split into stages: read the PDF drawing, identify items and quantities, price the materials, produce the quote. At each stage a person checks, corrects and approves the AI's draft before it moves on.

A dashboard app that transcribes a meeting in the browser as it happens, then has a lead AI dispatch work to specialist agents that assemble tasks, owners, decisions, agenda items and a summary diagram. The facilitator is freed from note-taking, and when the meeting ends an organised board is already on everyone's screen.

An agent for foreign visitors caught in an earthquake or typhoon in Japan, unable to read the warnings or tell where to evacuate. Drawing on official meteorological and geospatial data plus preparedness guides, it delivers alerts, shelter directions and ready-to-send safety messages in the user's own language.

Upload one photo of the furniture you want to build and this web app works out the parts, generates a 3D model, lays out the cutting and assembly steps, and exports a plan as a PDF. It targets DIY beginners who cannot draw plans and hesitate over materials, so they can start work already holding instructions they trust.

A phone and tablet recipe app built around the annoyances of cooking: messy hands, jargon you keep looking up, paper recipes you cannot use. A wake word starts voice control that answers instantly from a local model, and questions about terms or substitutions are handled with RAG. Photograph a magazine page and it becomes data.

An agent aimed at the paperwork load on social workers. It fills a client's assessment items from an interview transcript, then searches a custom RAG index of support programmes in Yamagata Prefecture and lists fitting ones as tasks. A plain UI works where paper and digital coexist, returning time to face-to-face support.

A web service linking the owner of a lost pet with whoever spots it. One agent reads images, another predicts behaviour, and together they estimate from weather, time of day and terrain where the animal is likely to be, drawing it as a heat map. Finders' photos are auto-matched by similarity, and both sides can chat directly.

A web app for marking handwritten exams. Questions and grading criteria are registered in advance; given a PDF of answers, three AI roles, sorting, transcription and commenting, draft the feedback. Teachers edit and place the comments on screen, then export a PDF, lifting a returned test from a bare score to real feedback.

A system that surfaces road hazards before you drive them. It pulls Street View imagery every 50 metres along a chosen route and has AI analyse each point from several angles — blind corners, road surface, visibility — scoring risk on a ten-point scale. The result is a colour-coded map you can question spot by spot in chat.

A physical device that turns a petty argument into a debate tournament. Each side speaks in turn; the speech is transcribed and an LLM moderates, scoring on logic and whether claims are backed up. The score travels over MCP to BLE motor control and appears as the tilt of real scales. Swap the prompt and the rules change.

An app for experiencing race day before it arrives. It cross-references the course GPX, the weather and the runner's goals and training history, then works out which points to watch, how to pace and where to drink. The advice comes back as a generated video with captions, helping runners picture the course and judge well.

By Texas grocer HEB: intelligently builds customer-specific shopping lists from events, lifestyle, budget and brand preferences, aligned with store inventory.

A natural-language interface for stock screening — converts user queries to screener API calls and refines filters through conversation.

For UK broadcaster ITV's streaming service ITVX: an AI chatbot that recommends content matched to each viewer's unique taste.

Turns any text or file into an immersive tutoring experience — personalized study guides, tests and context-driven sessions on Azure AI and GPT-4.

An educational history game combining immersive storytelling with interactive AI-driven dialogues.

Upload lecture notes and Insight transforms them into personalized quizzes and flashcards.

A classroom-integrated tool where GPT-4 turns student notes into tailored flashcards and quiz questions in real time.

A supply chain plan that drifts from reality can become a stockout before the analysis is done. eCoino keeps several response scenarios side by side in a BI view and lets planners work through them with a facilitator-style AI agent that drafts countermeasures quickly. Gathering is delegated; the decision stays human.

An end-to-end teaching assistant for new English teachers at tablet-equipped junior high schools. It spans lesson prep, the class itself, automatic quiz generation and grading, and per-student feedback. Classroom audio is transcribed so questions match what was taught, and wrong answers are analysed class-wide.

As an organisation grows, teams lose sight of each other's work and editing the internal newsletter turns into a chore. Here AI agents read company chat and meeting notes on a schedule, then lay the week out as a newspaper page, editorials and staff profiles included, so internal news becomes something people read for fun.

A phone app for travellers stuck in front of an unreadable menu. OCR pulls dish names and prices from a photo, then adds a translation, a short explanation and a generated image so you can picture the taste and look. Pick a dish and the app renders your order in the local language to show the waiter, and it keeps past lookups.

A weekend outing planner for families with small children. Register the area you live in and the agent keeps collecting local events from the web into a database, then uses vector search to surface the ones closest to your tastes. Pick one and it builds a full day around it, including travel, meals and places to change a nappy.

Pregnancy brings changes day by day, and memory alone makes them hard to describe at a check-up. BLOOMS is an iOS and web app that takes over the record-keeping: generative AI responds to the notes and photos a user leaves, suggests keywords to look things up, and summarises a stretch of entries to bring to an appointment.

A web app for households where two working parents and the price of picture books squeeze out bedtime stories. Pick an age, a script, a length and a genre, and the AI writes a story, illustrates it and reads it aloud. At around a dozen yen per episode, a fresh book can match the child's stage and interests every time.

An AI coding agent aimed at the cost and time a mobile MVP takes. Describe the app you want in plain language and it writes Flutter code you can preview on the spot, and when the build breaks it repairs its own errors. Finished apps ship to Firebase in one click, letting a solo developer cycle through prototypes quickly.

An AI chatbot for room rentals and hostels — personalized recommendations claimed to cut sales-department costs by an average 36%. Won a spot in the Slingshot accelerator.

A culture-fit bot that helps companies identify candidates aligned with their culture, work style and values.

Uses Whisper and GPT-3 to transcribe and analyze stock-trading videos and help implement the strategies they teach. Won among 1,700 participants in the early days of generative-AI hackathons.

Analyzes live microphone speech with machine learning to flag obscene, toxic, threatening or insulting language.

A chatbot for people who stutter: accurate speech-to-text from live voice recordings, plus answers to personal questions about stuttering.