AI Hackathon Winners of 2026

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.

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.

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

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.

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 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.

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 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 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.

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 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 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.

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.

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.

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.

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

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

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.

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

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

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

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

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.

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

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.

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.

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

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'.

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

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 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.

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.

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.

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.

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 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.

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.

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 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.

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.

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

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.

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 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%.

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.

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

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).

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.

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.

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

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.

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

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.

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.

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.