# Submissions: Gen AI Summit Hackathon

47 teams submitted a project. 9 reached the semifinals, and 4 pitched in the finals at the GenAI Summit.

## Winners

### Otty

- **Result:** 1st place
- **Track:** Open Track
- **Members:** Alex Koukos, Ioan Croitor Catargiu, Anemos Aloupis, Elena Paleocosta
- **Demo video:** https://youtu.be/wi6PpO3ClFg

Otty is a small and medium business's own AI agent. Paste the business's website, and Otty reads it and opens one door (an MCP server) that any customer's AI assistant can use to ask, order and book, inside the owner's rules. When something changes ("size 42 is sold out"), the owner says it in one sentence and Otty re-settles every affected order with each customer's assistant, asking the owner only when a rule would break. Every action and every refusal gets a signed, tamper-evident receipt.

### Steptrace

- **Result:** 2nd place
- **Track:** Open Track
- **Members:** Uipko Stikker, Ioannis Vossos, Arsen Tsntsgoukian, Nikolaos Taiganidis
- **Demo video:** https://youtu.be/itl3_OurC98

StepTrace turns any how-to video into an AR guide: point your phone at the thing you're fixing, and it highlights the exact part for each step.

### Flux

- **Result:** 3rd place
- **Track:** Open Track
- **Members:** Alexandros Nomikos, Ledi Brati, Nikos fourlis, Angelos Kokiasmenos, Grigoraki Dionysia
- **Demo video:** https://youtu.be/E-xKTwRW93E

an end to end translation tool for the deaf community. (sign language to speech and vice verca)

## Finalists

### FIREYE

- **Result:** Finalist
- **Track:** Open Track
- **Members:** Nikolas Sapalidis, Marco Serat
- **Demo video:** https://youtu.be/F1bdtWd3L4o

FIREYE is a wildfire command console for fire chiefs and incident commanders.

In the first hour of a fast-moving fire, information comes from everywhere at once: radio, cameras, drones, weather sensors, and news. It arrives late, out of order, with no indication of how old or reliable it is.

FIREYE brings everything together on a single 3D satellite map. Fourteen simulated camera sources (six drones, each with visible-light and thermal channels, plus eight fixed cameras) and 24 ground sensors share one view.

Click any feed to explore a 3D view of the terrain, with the fire reconstructed as it was at that moment.

An orchestrator sends drones to scout high-risk areas that fixed cameras cannot see, then brings them back to land and recharge. Every task shows its reason.

Alerts trigger only when something materially changes: a new fire, spread accelerating by 20% or more, a predicted arrival time moving earlier, or a location becoming newly threatened within the next hour.

Every piece of evidence displays its source, capture time, and whether it is simulated, so old data never looks fresh.

Consequential actions require human approval. The backend enforces that only the current human Chief can approve a task, approvals expire, and stale or duplicate inputs are rejected.

The entire incident can be replayed exactly as it was known at any moment.

The demo replays the 2018 Mati fire near Athens, from 14:00 to 22:00.

## Semifinalists

### never alone / Omega Wonderful

- **Result:** Semifinalist
- **Track:** Voice & Enterprise Agents · Wonderful
- **Members:** Vlasios Zacharis, Apostolos Grigoris, Vasileios Roumeliotis
- **Demo video:** https://youtu.be/mRTLPj5RQzg

Never Alone is the proactive AI care agentic system for older people. Older people leaving hospital after surgery, or living alone with a chronic condition, are on their own between visits. They don't call because they don't want to be a burden, and nobody checks on them. For each person we draft personalized instructions and a care plan based on their history, their medicines and their surgery, and their doctor or nurse approves it. Then we call them proactively and ask about their medicines, their instructions and how they feel. Every call, every alert and every change updates their personal risk profile and score, with the reasons. So the nurse sees who needs her first, the same day. When they press the emergency button, we call back immediately. Their family knows they're okay, and they stay at home, not back in hospital. We sell to private clinics and hospitals: they upload their patient list and go live in a day. Their nurses only see who needs them, fewer patients come back, and every patient leaves with someone who calls.

### Solon / Simeron

- **Result:** Semifinalist
- **Track:** Open Track
- **Members:** Gregory Sarantinos, Costas Sarantinos
- **Demo video:** https://youtu.be/zvJ07xPyDiU

Solon is a personal A2A & A2P financial agent in front of your money. Built for everyday people juggling bills, subscriptions, and taxes on a tight budget, with a Zero-PII A2A negotiation to rank what to pay first by cited law and uncover state rebates. A 5-gate deterministic firewall blocks scams, autopay duplicates, and budget breaches before 1-tap Passkey + AP2 SD-JWT approval, making agentic payments safe: models read, code decides, you approve.

### eleven0zero / aimilios & co

- **Result:** Semifinalist
- **Track:** Open Track
- **Members:** Aimilios Potoupnis
- **Demo video:** https://youtu.be/1RApGzV-77I

eleven0zero is a cybersecurity tool that watches for new vulnerabilities in real time. When one appears, AI agents take action. The first agent pinpoints which of your servers are exposed, proposes a fix, and takes preventive measures to shield critical infrastructure. The second tests the fix on a sandboxed clone. The third generates an exploit and runs it against the patched clone to prove the fix holds. It all takes seconds, and no testing ever touches production.
Everything is logged in a dashboard to be reviewed by IT engineers.

### ClinRadar

- **Result:** Semifinalist
- **Track:** Biotech & AI · Athens LifeTech Park
- **Members:** Ioannis Panaritis, Konstantina Imprisimi
- **Demo video:** https://youtu.be/ji6FRsEAoS4

**What it does:** ClinRadar prevents clinical supply disruptions. It watches the countries where your trials are active for anything that could stop a shipment: strikes and transport disruption, severe weather and natural hazards, security and travel advice, and changes to regulations and customs rules. Before dispatch, it checks every planned shipment against those events and against the destination country's requirements, such as import permits, labelling, importer of record and delivery days. A shipment that passes every check is cleared as ready to ship. If one will be disrupted, Clinical Supply gets recovery options that have already been checked against your rules, ready to approve.

**Who it helps:** Clinical Supply teams.

**Why it matters:** It prevents business disruptions so that patients' dosing is never compromised.

### Jarvis Assembler / NP-TRYHARDS

- **Result:** Semifinalist
- **Track:** Open Track
- **Members:** Giannis Karyotakis, Foivos Timotheos Proestakis, Gerasimos Athanasiadis, Vaggelis Talos
- **Demo video:** https://youtu.be/6LZG1k2-EN4

Jarvis Assembler turns the wordless paper guide of any flat-pack furniture into a 3D build you can't get wrong. You scan the box and an AI agent (Claude Agent SDK using CadQuery and Blender MCP) rebuilds every part and animates every step in minutes. Then pick up your phone or put on a VR headset: it watches you build through the camera, tracks the real boards on your floor, overlays the 3D guide on the furniture itself, and tells you by voice in real time whether the part in your hand is the right one or what the next step is.

## All other submissions

### Housekeeping Manager / The Housekeepers

- **Track:** Open Track
- **Members:** Kerkyra Dimisianou, Foteini Sfakianaki, Vaggelis Farmakis, Manousos Linardakis
- **Demo video:** https://youtu.be/UHstZ5xnI7Y

Housekeeping Manager is a cleaning-team scheduler for hotels and short-term rentals. The owner enters bookings in a web page or imports them from Excel, and the app creates the cleaning jobs and offers each one on Telegram to the best cleaner, based on distance and workload. Cleaners reply in plain Greek or English, such as "I can only do job 4, not the rest" or "I'm sick until Friday", and Gemini reads the message and re-plans: it moves the jobs to the next-best cleaner and alerts the owner in the dashboard if none is available. New team members join with a one-tap Telegram link sent by email and the owner follows everything live on the dashboard.

**Built during the hackathon:** We build the whole project during the weekend. We use Gemini API for answering to the cleaners and the Telegram bot for communication.

### CareerChat / Efthimios Charokopos

- **Track:** Open Track
- **Members:** Efthimios Charokopos
- **Demo video:** https://youtu.be/arQ3VP8tHPI

CareerChat: a personalized career guidance platform that help high students find the suitable direction and academic path.

**Built during the hackathon:** During the hackathon, I built CareerChat, a GenAI-powered career exploration prototype for students aged 15–18.

Students begin with a narrative-style conversation that identifies their interests, values, and preferred work environment. CareerChat then creates a starting career profile and introduces relevant professional avatars. Students can speak with these avatars, learn about their working lives, compare career possibilities, and complete realistic work-based activities.

For example, Jenny the Journalist guides students through “The Truth Detective,” an activity where they evaluate evidence before deciding what can be published responsibly.

The goal is to help students move beyond static career descriptions and experience what different professions may feel like before choosing a career or study path.

### Bid-Ready / solo levelling

- **Track:** Open Track
- **Members:** Nikolaos Kostopoulos
- **Demo video:** https://youtu.be/0C5STIg84WE

Bid-Ready is an agent that does the tedious, multi-step work of bidding for a Greek public contract, and checks itself before calling anything ready.

Who it helps: small Greek companies (5–50 people) that could deliver public contracts but never bid. 67.9% of Greek startups have never bid on one (Powergame, 2024).

What it does: from two lines about the company (or its website, via Firecrawl), it builds a profile, searches open EU tenders on TED, and ranks them with a visible fit formula. For a chosen tender it finds the tender document (including in the Greek ΚΗΜΔΗΣ register), reads every eligibility rule with Gemini, and answers ✓ / ✗ / ? per rule, each with its PDF page. It lists what you’d have to provide (guarantee, documents, deadline) and builds a handover pack: a to-do checklist, the tender, the receipts, a deadline calendar file, and a Greek letter to a consortium partner when you can’t bid alone.

Why it matters: one wrong figure gets a bid excluded, and LLMs confidently repeat wrong figures. So an independent verifier re-reads the cited PDF page for every figure and quote and writes a receipt; nothing is “ready” while a check has failed. In the demo it catches a real trap: TED’s €300,000 headline leads to a €6,000 guarantee, and the tender’s page 27 says €4,000. In our eval Gemini got 27/27 facts right when given the page, and fell for the headline every time it answered from it alone. Oumi adds a second opinion on each rule. The same tools are exposed over MCP, so other agents get a receipt, not a vibe.

**Built during the hackathon:** Before the official start (Sat 13:00) I prepared fixtures only, picked a real tender (TED 769674-2026), downloaded and extracted its documents, wrote a hand-made clause index for it (requirements.json, used to know what to check and where, never to grade), and confirmed the TED API works.

Everything else was built this weekend (31 commits, from Sat 16:03): the web app and API, onboarding (with Firecrawl website reading), TED search and scoring, the generic rule reader for any tender (Gemini), ΚΗΜΔΗΣ document lookup, the eligibility verdict with self-declared answers, the independent verifier and SQLite receipts, the Greek letter with a claim check, the handover pack, the Oumi second opinion, the Gemini eval harness, the MCP server (8 tools), 74 offline tests and the logo. The second demo tender (774438-2026) has no hand-made index: its rules are read live and verified.

### Vulisi

- **Track:** Open Track
- **Members:** Dr. Georgios Stathis, Panagiotis Kriparakos
- **Demo video:** https://youtu.be/-ecmBiN01P0

Vulisi captures what experts know and hands it to people and AI agents. It turns an interview into knowledge cards, finds the rules nobody said out loud in past decisions, lets a person confirm them, and tests the agent on decisions it never saw. In our synthetic demo, an agent went from 46 to 80 of 80 tricky invoice lines. For deployment strategists and anyone who relies on knowledge in someone else’s head. Free at app.vulisi.com, with files you own.

**Built during the hackathon:** Everything was built this weekend: the repo started empty on Saturday at 12:57 and has 165 commits since. New: the synthetic accounting dataset with hidden rules, capture, gap-finding, the held-out scoring, the web app (accounts, private libraries, own AI key), the website and the demo video. We built on existing tools only: Claude and Gemini through their APIs, FastAPI, Obsidian’s Markdown format, Google sign-in and Render hosting.

### Hireful / Team Beta

- **Track:** Voice & Enterprise Agents · Wonderful
- **Members:** Antonis Fourfaro, Oleksii Onufriiev, Thalis Theodoulidis, Yuriy Zhayvoronok
- **Demo video:** https://youtu.be/1TFYUm8Q0o4

An agentified recruiting platform that helps hiring managers/interviewers screen candidates and schedule interviews.

We offer options for intelligent agentic queries to find who might be the best fit for a position, as well as the option to use voice agents for initial screening calls.

The goal is to assist and accelerate the hiring process, saving a lot of time for enterprises with large workforces and significant turnover.

**Built during the hackathon:** This was entirely built during the hackathon on top of the wonderful platform.

### Roomify / Roomify Team

- **Track:** Content & Digital Art · fal
- **Members:** Dimos Vasalakis, Theofanis Papaspyrou, Stylianopoulos Panagiotis
- **Demo video:** https://youtu.be/-W4ENIuAbtg

We built Roomify, an AI agent that furnishes your room with real products that fit your budget.

You upload a photo of your room, or send one from your phone with a QR code. Then you describe what you want by typing or speaking. Roomify searches retailer listings, including Greek shop catalogues, and checks each product against your brief, your budget and the measurements it has. It then renders the chosen pieces into your photo. You get a furnished preview and a shopping list with real prices, dimensions and links taken from the retailers' own pages.

The agent works in the open. You can follow its progress live, see why it picked each product and see why it rejected the others.

You can keep changing the room after that. Add, remove, move or swap pieces with the item controls or in plain language, like "swap the sofa for something cheaper in grey." You can also compare the result with your original photo. To get a better sense of the space, you can explore an interactive 3D reconstruction of the room: orbit it, switch camera angles and inspect each product's model next to its listing photos and dimensions.

We built all of this in 36 hours. The goal was to close the gap between "I wonder how this would look" and "here's exactly what to buy."

### Michael

- **Track:** Open Track
- **Members:** Manolis Maris, Loukas Vetoulis
- **Demo video:** https://youtu.be/WhiJvZteRG0

Have you ever wondered what makes you lose time and deviate from your goals? We have. That is why we built Michael: an app that proactively saves you time and makes you more efficient. He is for founders and executives with very little time.

Michael is not a productivity app that just blocks your apps for two hours. He works continuously to analyse your behaviour, shows you exactly how you lose time and what breaks your efficiency, and acts proactively to help you fix it. Imagine a chief of staff who has watched every hour of your day. That is Michael.

What he does, and how:

1. He collects your data and biometrics through your existing apps and devices: Apple Watch, Google Calendar, Mac and iPhone app use, the browser, Claude Code and location. One timeline of your whole day. Nothing new to wear and nothing to fill in.

2. He stores them securely, in the EU, and analyses them to find the patterns that make you less efficient. One score: the share of your planned goal time you actually spent, with the minutes lost and what took them. Then a daily table of 34 figures, where every sensible pair is tested, same day and next day. He finds links you would never connect yourself, and chains of three such as late phone, lower sleep score, slower first block. Claude proposes further tests, the server verifies each one on the data, and the model never supplies a number. A pattern is kept only when it has held at least 4 times and passes a strict statistical test, so every result reads "held N of M times".

3. He looks at your real time data and lets you know, in three ways. A message when something is off, on your iPhone, your Apple Watch and Telegram, at most 5 a day. For example: "This week your phone opened every 14 minutes between 09:00 and 18:00. That cost you at least 3h 31 of focus this week." A recommendation on how to optimise your schedule. For example: "Leave at 19:15 instead of 18:30, win 20 minutes." And a proactive action, with one tap from you and with undo. People text you while you work, so he finds the hours your phone pulls hardest and offers Do Not Disturb for them, and it ends by itself. He also books a block for the goal that is behind, moves your own blocks, or holds a rule such as "Phone down at 23:45" every night.

4. What he cannot understand himself, you tell him. Each morning: what would make today a win. Each evening: what went right and what did not, in about a minute. He keeps it, counts it in your score, and uses it in his next advice and in the patterns he learns. He remembers.

5. He is connected to the apps you already use, and works with them for your benefit. Ask Michael anything about your day in the app, by text or out loud. Or connect Claude with one link and ask there: Claude reads your day, your goals, your score and what Michael has learnt, so the answer comes from your own data. Every agent query is logged, and you decide what each agent may see.

Why it matters to someone with no time to spare: Michael asks for about a minute a day and sends at most 5 messages. You do not study a dashboard. He tells you the one thing that is costing you time, and fixes what he can himself. So you spend your time on what matters.

Private by design: every source is opt in, Off the record stops all collection in one tap, and the database is in the EU.

### Palladion / Multi-Agent Therapy

- **Track:** Open Track
- **Members:** Spyridon Bantis
- **Demo video:** https://youtu.be/ARaYXjiwRSs

Palladion is an AI operations desk for wildfire evacuation and rescue. Messages from people in danger arrive from any channel, including a public web page anyone can open on their phone. A team of 15 Gemini agents on Vertex AI reads each one. The agents work out who and where the people are, check the fire’s path and the escape routes, judge credibility and medical needs, and catch duplicates. Code verifies every claim the agents make: each field must quote the message and each number must come from the fire and road models. Every case reaches the operator with a proposed decision and a deadline: the minutes until the front arrives. The operator approves or changes it with one click. Oumi Iris gives an independent second opinion, and nothing that moves people or units happens without a person. The desk also drafts 112 warnings, Coast Guard requests and crew briefs, and the sender sees the operator’s reply live on their phone. Any region or fire is a configuration file, not new code. We demonstrate it on the 2018 Mati fire in Greece, replayed minute by minute with real maps, wind and burned-area data. It runs hosted on boxd.

### Quantum Table / Q-Scientist

- **Track:** Open Track
- **Members:** Ioannis Theodonis
- **Demo video:** https://youtu.be/-U82UrLSq64

A quantum dice table run by a team of AI agents. Your dice are qubits entangled with the house’s dice, so neither you, the house nor an agent picks the number. Before the dice land, a verifier written in code (not a model) checks every roll against the exact quantum prediction. Every roll is then published as a public record, and anyone can recompute it with one command. It’s for students meeting quantum computing for the first time (they play first and learn entanglement along the way), and for game and AI-agent builders who need a random draw they can show they didn’t rig. It matters because every random draw today is a black box, and AI agents are now making those draws too.

### Quorum / Team Quorum

- **Track:** Biotech & AI · Athens LifeTech Park
- **Members:** George Apostolakis, Ilias Pantazis
- **Demo video:** https://youtu.be/XIakgunzl6Y

Quorum finds the patients who fit a clinical trial and shows the proof for every criterion. A trial site picks a recruiting trial from ClinicalTrials.gov and Quorum screens its whole patient list against it. Plain code first drops the patients who plainly cannot fit, at no cost. AI agents read the rest criterion by criterion, and every answer must quote both the trial and the patient's note. Code checks that both quotes really exist and does the numbers, units and dates itself, so an invented lab value is rejected. Where the record is silent the answer is "unknown", never a guess, and the coordinator answers each missing fact once for all patients. The doctor picks who to screen, and the referral is written as a draft in their own Gmail or Outlook. It helps trial coordinators who today read 20 to 40 rules per patient by hand, and patients who never hear about a trial they qualify for.

### Ergon / Sax & Raw

- **Track:** Open Track
- **Members:** George Tzimokas
- **Demo video:** https://youtu.be/yt13mQOfvy0

Ergon is the AI agent for large construction projects.

What it does: Every week the BIM model of a building changes. Ergon compares this week's door schedule (5,000 doors x 50 parameters) with last week's and with what the door manager asked for. It tells you what was done, what is wrong, what is still missing, and what changed that nobody asked for. Safety changes (fire rating, smoke doors, panic bars) are flagged as LIFE SAFETY – STOP. It writes an Excel report and a CSV to update Revit, sends each team only its own findings (Telegram, email, or their own AI agent via A2A/MCP), runs by itself every Monday, and keeps a record of every week: what was checked, what was found, and who was told.

Who it helps: the people who coordinate a building's model every week (door, window and equipment managers, BIM coordinators) and the teams that depend on them: architects, fire safety, suppliers.

Why it matters: Today this check is done with Excel formulas and takes a full day every week, and mistakes slip through. A fire door downgraded by accident is a real safety risk. A chatbot can't do it: the data is too big and it guesses. In Ergon the code does the exact counting, and AI only explains it, so the numbers are always right. Today it checks doors; tomorrow windows, steel and timber: the same engine, with new rules.

### Regula / +One

- **Track:** Open Track
- **Members:** Margarita Varla, Aggelos Ladas, Maria Mitropoulou, Raldi Komini, Athina Feleki
- **Demo video:** https://youtu.be/QJm0Ef44s1E

REGULA is an AI-powered compliance office that helps companies understand exactly where they stand on EU AI Act, GDPR, and AI governance requirements—in minutes instead of weeks, saving money and reducing the risk of unexpected fines.

Powered by 18 specialized AI agents, REGULA scans a company's website using a real browser, gathers essential information through a short discovery process, and evaluates 28 legal obligations. Every finding includes the relevant law, supporting evidence, and recommended actions, with human oversight for final approval.

REGULA also generates regulator-style inspection letters, compliance reports in Word and Excel, and data sovereignty maps showing who can access company data. Through its MCP integration, users can even query their compliance findings directly from Claude or ChatGPT.

From “Does this regulation apply to us?” to a clear, actionable compliance plan in a single afternoon.

REGULA is a compliance support tool, not a substitute for legal advice.

### Homie the agent that gets you the keys / Dimitris Tzouvaras

- **Track:** Open Track
- **Members:** Dimitrios Tzouvaras
- **Demo video:** https://youtu.be/uI5S2gCsMpc

Homie is an AI agent that finds rental homes and makes the phone calls for you.

What it does. You type the home you want. Homie shows its plan before it does anything, reads the real listings onto one board, looks at the photos, and marks the homes worth a call. You pick the homes. A voice agent then calls the landlord, says in its first sentence that it is an AI assistant and who it is calling for, asks what matters to you, and holds a viewing at a time your calendar is free. The viewing is written to your Google Calendar and the landlord gets a text with your details. On the same board a crew of eight agents (coordinator, scout, photo inspector, caller, receptionist, skeptic, scheduler, analyst) works through gated tools, and the coordinator decides the next step after every call. Any MCP client, such as Claude, can plan a search and follow the board.

Who it helps. Renters who cannot spend the working day on the phone. People who do not speak Greek: Homie calls in Greek and shows the conversation in English, line by line. People who are deaf, hard of hearing or cannot speak on a call: the whole call is text on the screen.

Why it matters. Agents stop where the other side only has a phone. In Athens the listing is online but the landlord is a phone number, so finding a home still means dozens of calls: no answer, already rented, no pets, call back after six. Homie does that job end to end and comes back with a booked viewing.

### MnaMoN

- **Track:** Open Track
- **Members:** Exarchos Theodoros, Konstantinos Bouros
- **Demo video:** https://youtu.be/EjzqUlwbrdg

What it does: Mnamon is an AI agent that audits AI systems against the EU AI Act. Given a system and its model, it works out which legal requirements apply, plans test scenarios, generates synthetic data, tests the real model in a sandbox, and reports findings. Each finding cites its legal source and includes a critic's counter-arguments. Humans approve the plan before anything runs, and the agent remembers rejected tests. When the law changes, it finds the affected systems, re-runs only what's needed and shows what changed.

Who it helps: AI engineers and compliance teams at companies that build high-risk AI, such as hiring, credit scoring or benefits decisions.

Why it matters: The AI Act requires these companies to prove their AI is tested, fair and kept up to date with the law, with fines up to €15M or 3% of turnover. Today that proof is manual, slow (takes several months with back and forth) and out of date when guidance changes. Mnamon makes it automatic, repeatable and auditable. Every audit is pinned to the exact model and law version and traced step by step, and it runs in-house, so models and data never leave the company.

### VoxEye / VoxEye team

- **Track:** Content & Digital Art · fal
- **Members:** KAIPIS MICHAIL, DIMITROPOULOS KONSTANTINOS, ALEXANDROPOULOU KONSTANTINA
- **Demo video:** https://youtu.be/sIo0kuK4dTA

We are the VoxEye team (Konstantina Alexandropoulou, Michail Kaipis, Konstantinos Dimitropoulos) and we created VoxEye — an AI-powered voice assistant that brings independence and safety to people with visual impairments.

Through natural, real-time voice interaction, the app turns a smartphone into a pair of "eyes":

Smart Walk: Navigates city streets and describes obstacles or points of interest.

MedScan AI: Accurately reads medication packaging, dosages, and labels.

Instant SOS: Automatically shares the user's live location with emergency contacts.

This is only the beginning — in the near future, VoxEye is set to evolve even further by integrating advanced computer vision capabilities, offline support, and smart wearable connectivity, making everyday life truly accessible for everyone

### UNSLOP / The ripple effect

- **Track:** Content & Digital Art · fal
- **Members:** ΕΥΣΤΑΘΙΟΣ ΠΑΡΑΣΚΕΥΟΠΟΥΛΟΣ, ΓΕΩΡΓΙΟΣ ΒΑΜΒΟΥΡΑΣ
- **Demo video:** https://youtu.be/-VL1LvBJuhw

Cheap AI ads are everywhere now, and people scroll straight past them. Small businesses still can't pay an agency to do better and agencies aren't even caught up. UNSLOP gives them the research an agency would do first: it reads their site, asks them a few questions, and sends agents out to real customer reviews and the ads their competitors are running. From there, it follows a 13-step strategy built on more than fifteen hundred real ads and on pharma market access practice, where every claim needs evidence. The owner picks statics, video or both, and gets the ads, a competitor report and a marketing plan. It works on the website, or inside Claude through our MCP connector on the owner's own subscription.

### T-Minus / Perceptrons

- **Track:** Open Track
- **Members:** Stavros Peppas, Jabran Raja
- **Demo video:** https://youtu.be/NaiGAQudO0U

You don't know when an agent finishes his work. But now with T-Minus you do, you can get an estimate if a task is feasible or not. easy and fast. It gets the workdone in T-

### Merc / Team Delta

- **Track:** Voice & Enterprise Agents · Wonderful
- **Members:** Filippos Papageorgiou, Maria Delaporta, Klavdios Sklivanitis, Thanasis Maragkos, Stefanos Kampouris
- **Demo video:** https://youtu.be/9a-YpRJU0Cg

Field Service Sales Personal Assistant with an AI Native CRM

Merc: a CRM that fills itself. Reps talk to a voice buddy about each visit, and Merc logs it, briefs them before the next meeting and follows up after.

### ProtocolPulse / LowBall

- **Track:** Biotech & AI · Athens LifeTech Park
- **Members:** Stefanos Theos, Athanasios Bitsanis, Emmanouil Sanoudakis
- **Demo video:** https://youtu.be/Fgcc1lhQZOM

ProtocolPulse is an AI-assisted platform that helps clinical teams and trial sponsors move from protocol to action. It extracts inclusion and exclusion criteria into source-linked rules for human approval, screens patients, flags missing evidence and monitors new clinical records for protocol conflicts. AI investigates findings and drafts cited summaries, while clinicians retain control of decisions. By reducing repetitive review and keeping evidence traceable, ProtocolPulse helps teams run more efficient trials—with the potential to bring new medicines to patients sooner.

### AIz

- **Track:** Open Track
- **Members:** Αχιλλέας Βιλλιώτης, Βασίλειος Μπαρδάκης, Χαράλαμπος - Μάριος Χάλλας
- **Demo video:** https://youtu.be/g5LhC7qgd1E

AIz is an adaptive assistant that observes repetitive workflows as you work and automates them on the fly. Powered by Gemini and Oumi Iris, it supervises the user while they work, and suggests improved workflows. For example, it can take over the task entirely and complete it in a single click, or suggest a shortcut to a user for something they use regularly.

### Once / Never Twice

- **Track:** Voice & Enterprise Agents · Wonderful
- **Members:** Michail Dadopoulos
- **Demo video:** https://youtu.be/i_GpY1wijJc

Once is an AI purchasing assistant that asks you once, then knows.

It helps people like Nora, who runs a small jam kitchen and does all the buying herself. Today, letting an AI agent spend money means picking between two bad options: an assistant that interrupts you about every step until you stop listening, or one that never asks and leaves you to find the mistake when the wrong order arrives.

Once is the third kind. It handles a purchase request end to end: reads it, compares supplier quotes, checks the budget and prepares the order. Before anything that can't be undone, it phones you, says why it chose this option, and waits for a spoken yes. The same ask arrives in Telegram with the reasons and a link to the evidence. Once it has done a job right, you can save those steps so it does them the same way next time, and teach it a rule in one sentence, such as "a replacement within 5% doesn't need me." After that the phone stays silent for what you've covered, and it still calls for anything new or outside your rule.

It matters because trust is what keeps businesses from letting agents touch money. Once lets that trust be earned one decision at a time, the way you would with a new hire

### gameHome / liapisBrother

- **Track:** Open Track
- **Members:** Evryviadis Liapis, Nikos Liapis
- **Demo video:** https://youtu.be/1AOXTsjMk9I

gameHome turns what a home listing already has, a floor plan and ordinary photos, into a walkable 3D tour. No scanning hardware, no operator, no site visit. A vision model reads the floor plan into rooms, walls, doors and windows at their real size. Each photo is matched to its room, and its camera position is solved from the image itself (vanishing points and edge alignment), so the real photo is projected back onto the 3D room exactly where it was taken. Rooms are then furnished in 3D from their photos. Owners upload their files or just paste a Prosperty listing link; renters and buyers get a dollhouse view and a first-person walk-through at true scale. It helps people judge a home's layout and size before visiting, and lets any listing with a floor plan offer a 3D tour.

### Shelf-It / Shelf-It team

- **Track:** Open Track
- **Members:** Anastasios Ntoukas, Ioannis Gkougkousis, Athanasios Karaferis, Christos Grigorios Psychogyios, Giannis Koutsourelis
- **Demo video:** https://youtu.be/KURUb9P58Bk

Shelf-it is an autonomous shelf auditing and merchandising system for grocery retail teams. Point a mounted phone camera at an aisle, and Shelf-it reads the shelf, checks every product against the store's planogram and backroom inventory, and generates a photorealistic preview of how that exact shelf should look once restocked and tidied.

Who it helps:
- Store clerks get prioritized restock and relocation tasks on Telegram with a cropped photo of the exact gap and one-tap claim buttons. They never tick checkboxes to close work: the camera closes each task automatically once it sees the shelf fixed.
- Store and category managers get automatic warehouse orders within configurable euro and case limits, layout swaps backed by peer-reviewed retail research, cited market evidence, and ready-to-print PNG price talkers.

Why it matters:
Across grocery retail, 8.3% of SKUs are out of stock at any moment. In 40% of stockouts, the shopper buys elsewhere or skips the purchase, even though 25% of "missing" items are already sitting in the store's own backroom (Gruen, Corsten & Bharadwaj, 2002). Manual aisle walks are too slow to catch them, while raw LLMs hallucinate arithmetic. Shelf-it separates perception from math: Gemini reads the image, while Python and DuckDB SQL compute every unit shortage, order cost, and margin gain with a full audit trail.

### Launch Kit

- **Track:** Content & Digital Art · fal
- **Members:** Dimitris Kyriakakis, Dimitris Kargatzis
- **Demo video:** https://youtu.be/XnsSUKJrEmc

Launch Kit turns one product link into a full launch. Paste a website, App Store, Google Play or GitHub link. It reads the live page for the brand's colours, fonts and logo, scans the competitors, and writes a brand brief, launch copy for nine channels and a 7-day social calendar, with every claim checked against the page. fal (FLUX) paints the backgrounds in one of three art directions you pick, the stills are typeset in five platform sizes, and HyperFrames cuts a teaser. A FastAPI service records a narrated demo video of the live product, optionally made for one prospect with an outreach email to send it.

It's for indie makers and small teams who ship an app and then have no time to launch it. A full kit takes about a minute.

### Recoup / All-Inners

- **Track:** Open Track
- **Members:** Odysseas Avdikos, Orestis Spyrou, Efthymis Grizanitis
- **Demo video:** https://youtu.be/5YAQ-J_3AiQ

Recoup is a safety-first enterprise agent runtime built for CFOs, financial controllers, and Accounts Payable (AP) / procurement teams to autonomously audit vendor deliveries and recover lost revenue without risking unverified mutations to production ERP ledgers.

Enterprise procurement teams lose 1–4% of spend to shipments that arrive short, damaged on the loading dock, past delivery SLAs, or billed above contracted rates. Yet enterprises block autonomous AI agents because a single hallucinated decimal, prompt injection in a vendor delivery note, or premature supplier email creates real financial and tax liability.

Recoup makes autonomous enterprise execution trustworthy through five architectural primitives:
1. Copy-on-Write ERP & Workspace Rehearsal (Safe Clone): Every SQLite ledger mutation, audit dossier, Greek AADE myDATA 5.1 tax credit XML, and outbound supplier email executes inside an isolated disposable clone first, producing row-level database diffs, unified file diffs, and SHA-256 drift verification before production is touched.
2. Self-Healing 7-Layer AST Tool Forge: Instead of trusting LLM prose for compound contract math (unit overcharges, shortages, damaged pallets, capped SLA late penalties), Recoup synthesizes pure Python calculators, validates their AST and OS resource limits in an isolated subprocess sandbox, runs TDD boundary tests with automatic self-repair, and registers verified tools live.
3. Zero-Output-Token System-1 Governance (Oumi Iris) + Deterministic Policy Guard: Hard Python rules (>= EUR 250 or external side effects) pause execution, while Oumi Iris evaluates parallel logprob heads (needs_human_gate, escalation_route, financial_impact) in 0 output tokens.
4. Live Voice Executive Consultation & Mobile Handoff: Stages dual settlement scenarios (full cash credit vs. physical redelivery) in the outbox and opens a live xAI Grok Voice + fal.ai Wizper STT consultation across desktop and /mobile iPhone view, using Oumi Iris to separate exploratory questions from binding spoken authorization.
5. Transactional Replay, Proactive AP Sweep & 1-Click LIFO Undo: Verifies SHA-256 workspace fingerprints against concurrent drift before commit, captures pre-image receipts for instant 1-click LIFO rollback, guarantees zero false claims on clean control shipments, and scans historical paid invoices in idle time to pre-rehearse unclaimed SLA recoveries.

### Hestia / NTUAI

- **Track:** Open Track
- **Members:** Yassin Ahmed, Despoina Christina Markatou, Vassilis Vassilopoulos
- **Demo video:** https://youtu.be/vn57sqhXk0U

Hestia is an autonomous AI housing agent that finds, ranks, and verifies homes from multiple sources. It uses Oumi’s Iris model to flag potential scams, then contacts landlords, checks your calendar, and books viewings through email or voice.

### xeAInhothika / Triton Dev

- **Track:** Open Track
- **Members:** Ioannis Argyras, Stavros Theofanopoulos, Manolis Manolis
- **Demo video:** https://youtu.be/OyAsJ-2bxpk

xeAInhothika is an AI partner for a 24-hour mini market, where one owner also orders from the suppliers, keeps the stock and does the books. A camera watches the shelf, deliveries are scanned by barcode, and invoices that arrive by email are read and checked against what actually came in and against myDATA, in the AADE test environment. It works out what needs doing and does the safe parts on its own (21 of 22 tasks in the demo day). Before anything that costs money (an order, a credit request, a new IBAN) it asks the owner on Telegram. Totals, VAT numbers and IBANs are checked by code, not by the model.

### myField / TokenMaxxers

- **Track:** Open Track
- **Members:** Panagiotis Kontoeidis, Stelios Dimitriadis
- **Demo video:** https://youtu.be/Bzj0cB_lV50

Every spring, Greek municipalities must verify property owners' declarations that they have cleared dry vegetation from their land. This is an explicitly manual process, but thousands of properties and limited inspection teams make it impossible to check everything before fire season. Consequently, uncleared plots can remain potential sources of wildfires, threatening nearby communities and forests.

myField automates the initial screening using satellite imagery, identifying potentially uncleared plots and presenting visual evidence. Municipal inspectors still perform verification and make the final decisions. 

By helping local authorities direct their limited resources where they matter most, myField aims to prevent wildfires, protect forests, and safeguard communities.

### DemoCut

- **Track:** Content & Digital Art · fal
- **Members:** Thanos Ziagakis, Vasilios Parikoglou, Eleutheria Kontou, Leon Long Nikolopoulos Tang
- **Demo video:** https://youtu.be/zMXMtmlD1lw

DemoCut turns a rough screen recording of your app, its GitHub repo and the judging brief into a short launch video, with nobody editing it by hand.

It reads the repo to learn what the app does and the steps a user goes through, watches the recording to find the moments that show each step, writes a script, and cuts the video from your real footage with voice-over, captions, music and title cards. Where the recording is weak or misses a step, it fills the gap with an illustration built from your own screens, clearly labelled as one.

What makes it different is the truth check. Every factual line in the voice-over is checked against the footage or the repo before it is spoken, and lines that can't be backed are rewritten or cut. In our submission video it checked 7 claims: 4 were backed, 2 rewritten and 1 rejected. It even caught its own model inventing a player's name that wasn't on screen. A final rubric of seven checks then judges the finished cut.

It helps anyone who has built something and has to show it: hackathon teams in their last hour, founders shipping a launch, open-source maintainers cutting a release. The demo video is the part of the project most people see and the part most teams have the least time for. DemoCut makes it in about eight minutes from a recording you already have, and it won't say anything about your product that isn't true.

It runs from a phone or laptop in the browser, and as an MCP tool that an AI agent can call.

### Pathien Health / onedotatatime.design

- **Track:** Biotech & AI · Athens LifeTech Park
- **Members:** Lefteris Georgaras
- **Demo video:** https://youtu.be/Do00XcMga-U

Problem
* Patients ask Google.
* They dontt tell everything about their med history
* Doctors decide from half a history and half exams or do assumptions

Solution
* User takes photo and digitize his med exams
* We translate to structural data and keep them secure
* Doctors can have access to the data after user consent
* We analyse the data so the doctor can think 10x faster to find a disease
* User can have the data all the time with him

Monetisation
* We can target users for clinical trials if they give concent
* We can reduce the yearly health budget up to 17,5%

### Brooking.ai / Gamma

- **Track:** Voice & Enterprise Agents · Wonderful
- **Members:** Georgios Alexakis, Vasilis Papadimitriou, Vasilis Bogdanis, Anastasios Papageorgiou
- **Demo video:** https://youtu.be/4t79EjxofLg

We’re building the system of intelligence for travel: one fleet of AI agents that connects travellers with the rental companies and hotels that serve them.

Today, rental companies run on fleet software and spreadsheets, hotels on their own systems, and none of them talk to each other. Every change is passed along by phone and email, one call at a time. At 11 PM on a mountain road in Crete, that’s far too slow.

We run a rental company ourselves, so we built both sides on Wonderful:

Fleetpilot is the operating system for rental companies: one place where the whole business runs, with agents that handle the calls, incidents and pricing around the clock, so the company sells more and its staff stop firefighting.

Brooking is a personal AI travel agent: it knows each traveller’s trips, family and preferences, and books and fixes everything in one conversation, so travelling feels like having a concierge on call.

### RollOut / Stellarfirm.ai

- **Track:** Content & Digital Art · fal
- **Members:** Nikandros Mavroudakis, Vasileios Drosatos, Ioannis Bilitsos, Odysseas Agiostratitis
- **Demo video:** https://youtu.be/8QgFml9EpvY

What it does
Rollout turns a live web app link or an Android APK into polished product videos. AI agents explore the app, write a script for each key feature, and record a real browser or emulator playing it. The recording is then edited automatically, with zooms, click highlights and cuts placed from what actually happened on screen rather than guessed from pixels. You get feature demo walkthroughs and short promo clips in one editor, ready to export. It also plugs into your CI/CD: every time you push to main, Rollout checks what changed and, if users have something new to see, makes the feature video for you automatically.

Who it helps
Founders, indie makers and small product teams and even big enterprises who build faster than they can make videos: startups, hackathon teams, and developers who need to show their product on their site, in docs, to investors or on social media, without a video editor or a spare day for recording.

Why it matters
Early-stage products change every week. Each time the UI or a feature changes, the old videos are out of date. Re-recording and re-editing them takes hours that a small team should be spending on the product. So most teams either show outdated videos or none at all, and people who can’t see a product working rarely try it. Rollout needs only a link, with no repository, backend access or editing skills, so it can remake accurate videos in minutes whenever the product changes. It can even do this automatically on every merge to main.

### Faceful

- **Track:** Content & Digital Art · fal
- **Members:** Dimitris Theodoropoulos, Minas Kontis
- **Demo video:** https://youtu.be/rn27D18xy8A

Faceful is a video studio that you and your AI agent share. Small businesses and agencies need short presenter-led videos (Reels, Shorts, TikToks), but making them is slow: a brief, a script, a presenter, edits, approvals and a posting plan.

In Faceful you describe a video and get editable, versioned drafts. You refine them by chatting with an AI Director, which proposes each change as a diff you can apply or undo. Your own agent (Claude Code, Codex or any MCP client) can work in the same workspace through a remote MCP server. It follows the same rules as people do, and it can propose changes but can never approve, publish or touch billing.

Every draft starts from a brand kit (voice, guardrails, products, presenters). Content is reviewed before it's scheduled, and nothing posts until a person approves it. The studio and agents share credits, billed through Stripe (sandbox).

It helps small teams who already use AI agents but have nowhere to put the work. An agent alone has no brand memory, no versions, no approval step and no billing. Faceful gives that work a home while keeping a human in charge of what gets published.

### PairShake

- **Track:** Open Track
- **Members:** Gregory Serderidis, Christoforos Koutsodimos, Jason shaw
- **Demo video:** https://youtu.be/DoCssfNAvHE

PairShake is an AI agent that finds you the right person to meet when time is short.

Whenever many people come together in one place for a few hours, at a conference, a hackathon, a meetup or an expo, the person you most needed to meet is often in the room and you never find them. There is no time to talk to everyone, and every conversation spent on the wrong person costs value.

With PairShake, your AI agent talks to the agents of everyone else in the room and finds the one person you should meet right now, and why. You both get an introduction card. If you both say yes, your phones guide you to each other across the room.

Fewer random conversations. More of the right ones.

Αν το portal έχει όριο χαρακτήρων, αυτή είναι μια σύντομη εκδοχή:

PairShake is an AI agent for crowded rooms with little time. When many people meet in one place for a few hours, your agent talks to everyone else's agents, finds the one person you should meet right now, and your phones guide you to each other.

### Metis · Agentic Lab OS / Metis

- **Track:** Biotech & AI · Athens LifeTech Park
- **Members:** Ertval Karameta
- **Demo video:** https://youtu.be/78atllOkIdo

Metis is the night shift for clinical-trial teams at biotech companies, CROs and hospital trial units. Trials stall: their eligibility rules shut out the patients they need, and problems are found too late. Every night, on her own, Metis gathers the data she needs (ClinicalTrials.gov, the CDC NHANES survey, drug labels), tests each trial's rules in disposable isolated copies, checks her own work and drafts the fix with its evidence. In the morning a clinician approves, holds or rejects; every decision writes a hash-chained receipt, and Metis cannot approve by design. She covers all four ALTP problem areas: protocol design and recruitment on real data (one diabetes trial: dropping one rule takes eligible people from 85 to 208, and women from 31.8% to 52.5%), patient experience on labelled synthetic diaries in Greek and English graded on PRO-CTCAE, and trial monitoring on the real registry (7 of 15 Greek recruiting studies overdue) plus a synthetic site study. People use her in a live operations console; other agents use the same core through the `metis` MCP server and CLI. The vision: Metis, the Agentic Lab OS.

### Insurance OS / Team Alpha

- **Track:** Voice & Enterprise Agents · Wonderful
- **Members:** Aris Bourounis, Apostolos Dedeloudis, Alexandros Melis, Panagiotis Georgakopoulos, Giorgos Savaidis
- **Demo video:** https://youtu.be/YV6ray2M2Gk

agent first company

### Hyper Health / Hyperminds

- **Track:** Biotech & AI · Athens LifeTech Park
- **Members:** Alexandros Bakagiannis, Jim Tsipoutas
- **Demo video:** https://youtu.be/kFbFm8oi00Q

Hyper Health is an AI layer between the patient and the doctor. Today, patients like Maria (68, diabetic, three doctors) carry a folder of paper and email the same PDFs to every doctor, and each visit starts from zero. With Hyper Health, the patient uploads her files once and speaks in Greek for 90 seconds. AI agents read her exams, connect them across sources and check them against clinical guidelines and drug interactions, with a source for every item. The doctor gets a one-page brief before the visit and confirms it in under 2 minutes. The AI prepares. The doctor decides. It never diagnoses.

### Prova

- **Track:** Open Track
- **Members:** Panagiotis Vardalas, Andreas Rigoutsos, Mike Fratzeskakis, Christopher Nyirenda, Elena
- **Demo video:** https://youtu.be/eAVlYPGEU4M

We built Prova, an AI call desk that lets a small insurance agency renew its policies by phone without breaking its own rules. You hand it the agency's messy files, such as Excel sheets, an old CRM export, a typo-filled price list, conflicting emails and WhatsApp notes. It cleans the data and writes its own rulebook, tying every rule to the sentence it came from. It ignores instructions hidden in documents and asks the owner only what it can't work out. Before going live, it certifies itself by attacking its own caller and deliberately breaking its own rules to check the tests catch it. Only the human owner can switch it on, with one tap. It then calls each client in their language, rehearsing three times against a simulated version of the client first. Discounts above the limit are refused by code, every agreement gets a written confirmation link, and sensitive cases like bereavement or complaints go to a human. It runs through a web dashboard for the owner or through an MCP connector, so an AI agent such as Claude can do the whole setup.

[Judging](https://hackathon.polymath42.com/judging)
