Today's projects
Everything the room built today. Pitches start at 15:15.
- #01formTrust, but check
Wasn´t Me
User is able to send sensitive documents safely to LLM for analysis without exposing the sensitive information. Application that takes in a document with sensitive information (e.g. medical report, invoice). The text is extracted, the sensitive information is detected and pseudonymized, and then the pseudonymized text is sent to the LLM for analysis. The user has to be able to confirm the pseudonymization process beofore it is sent to the LLM for analysis.
Wasn´t Me - #02via MCPWildcard
Who are you with AI?
A three minute, game-like run that tells you what kind of AI worker you are today, then opens a board game map of small experiments for your own job. Built for the colleague who has never opened a chat box: no scores, no ranking, no percentages, no product names, and it asks about your actual work before it says anything about AI. Two Claude calls rewrite every question around the real tasks of your week. Phase 1 is fully working and live; Phase 2 is one hand crafted example map, labelled as an example in the product.
Team DoraVanilla HTML/CSS/JS with ES modules, no framework and no build step, Claude Sonnet 5 called from the browser with server side web search, Cloudflare Pages, built with Claude Code - #03formWildcard
Dream Catcher
Every planning tool on the market starts at "enter your goal," which quietly excludes the majority of people who don't yet know what they want. Pink Carrot starts one step earlier: a psychologist-designed discovery that gets someone to a dream in their own words, then breaks it into concrete, costed steps. We sit at the exact moment a vague someday turns into a real date and a real price, which is where booking and saving both convert.
Team Carrots 🥕 - #04formTrust, but check
GreenRoom
AI-assisted scoring for conference programme committees.
47AIClaude API, VS Code - #05formTrust, but check
Devil’s Advocate: A News Checker That Argues Back
Devil’s Advocate turns news verification into a debate. It extracts factual claims from an article, searches for supporting evidence, and has an independent critic challenge the first verdict with its own sources. The result is a transparent trust score showing what is verified, disputed, or uncertain and why. It directly tackles the “trust, but check” problem: making verification faster than re-reading and researching everything yourself. It reflect the thread called "Trust, but check"
Trust NobodyFastAPI, Python, Claude API , ADK a2a, Haiku 4.5, Sonnet 5, Opus 5, Claude native web_search, Tavily Search API, Pydantic, SQLite - #06formWildcard
Journalist Assistant
The starting point is a six-agent design: Retriever, Fact-Checker, Style Agent, Planner, Publishing Support, and Orchestrator. It is not the number of agents that makes the system agent-based—it is the feedback loops: The Fact-Checker can force the Retriever to conduct further research; the Planner can change the format if the sources are weak; the Style-Agent must present the same content differently for the website, magazine, and podcast; and the Orchestrator must ensure that all of this complies with editorial guidelines. This document adopts the initial design and adds the elements needed for productive use in a newsroom. The first use case is the daily work of a health newsroom—that is where the specific need lies, and the technical examples in this document are based on that. However, the assistant should not be limited to this field but should remain adaptable to any use case where comprehensive and accurate research is essential; Section 4.7 addresses the architectural implications of this dual requirement.
Journalist-Assistant - #07formTrust, but check
Source Verifier — does the citation actually say that?
A lightweight demo checks links in pasted content, navigates to the linked pages to identify the source sentence, and displays the result as red, amber, or green. If red, it will suggest a better one. Also a Chrome extension allows one-click verification via right-click.
Börsen AmpelClaude API, python, fast API, JS, Chrome extension, Github pages - #08formWildcard
A radar tool that finds where agri PV reaches the grid fastest and realizes it full potential
AgriRadar; a tool that finds Agri-PV opportunities hiding in Germany's existing grid connections. We pulled the national biogas registry (MaStR) for Bavaria, grouped ~thousands of registered units into sites, cross-referenced them against solar installations to filter out farms that already have PV, and ranked the rest by installed capacity - since Germany's biogas plants were mostly oversized 2–6× during flexibilization and now sit idle ~75% of the time. The result is a ranked, mapped shortlist of real biogas farms with a documented reason to have spare grid headroom, plus an indicative Agri-PV revenue estimate per site - letting developers target sites with capacity that already exists instead of waiting years in Germany's 700+ GW connection queue.
AgriRadar - #09formWho gets left behind
Safe the Mittelstand
We built a proactive, personalized AI assistant for the sales reps of the German Mittelstand. In this environment, 70% of sales reps are between 40 and 60 years old. Many feel distant from their digital devices and are reluctant to hand over control. Our hypothesis: AI adoption is blocked by two things, the mode of interaction (typing and prompting) and the disruption to how reps already work. So we made voice the primary way to interact. The assistant empowers the rep by offering clear choices to decide on, and it takes over the tedious work like documentation, freeing them for the job that matters: selling.
SchraubiClaude API, ElevenLabs, GitHub - #10formTrust, but check
Check26
A checking layer around the summarizer. Every claim in the summary gets verified against the source by a second model in a clean session. Unverified claims are marked, not hidden.
Check26 - #11via MCPTrust, but check
Glassbox — does the research use its sources honestly?
Glassbox verifies AI-generated research as a separate loop from the original research — you bring the output, it audits it independently. It scores each source for credibility and relevance separately, then checks every claim against what its source actually says — supported, partial, unsupported, or contradicted — with a second adversarial reviewer. Live backend plus a 3D pipeline visualisation. Four recorded runs included: one scored 59/100 check_flagged, another exposed a calibration bug we then fixed. It ends by handing you a prompt to re-run the research better.
GlassboxPython 3.14, FastAPI, Anthropic SDK (Opus 4.6), httpx, pydantic v2, Three.js - #12formWho gets left behind
atina
Atina is a personalized learning experience designed to help people stay focused and learn more effectively. Instead of presenting everyone with the same linear course, it adapts the learning journey around the learner’s interests, knowledge, and learning type. For the MVP, we’re testing this approach with history. A learner can start with a broad topic such as WWII or Art History and explore it through connected events, people, ideas, and questions.
team atinaClaude Code, Next.js, Prisma, Notion - #13via MCPWildcard
From "what's for dinner?" to a filled basket
A meal-planning feature inside an online grocery app: it learns a household's allergies, tastes and pantry, plans meals everyone will actually eat, and turns the plan into a basket with only what's missing. Built for families who lose time and energy every week deciding what to cook. Working end to end — real backend, conversational planner and basket hand-off, covered by browser tests.
Off My PlateClaude API (Haiku 4.5 fast path, Opus 5 + Sonnet 5 multi-agent path), Node + Express, Vite, Capacitor, Playwright - #14formWildcard
Claude is a genius with amnesia. We fixed that.
llm-wiki — a disk for Claude's brain. Karpathy's LLM OS says the model is the CPU and the context window is RAM — wiped every chat. We built the missing long-term memory: Ingest turns our docs into neurons, Query answers from our knowledge with sources, Lint flags rot before it spreads, Draft writes what it learns back in. A wiki that maintains itself. Which thread does it pull? The one that unravels everything: if Claude remembers, everything changes. No more Groundhog Day chats. No more confident guessing. No more wikis that rot. And every agent we build next inherits the brain for free — because agents are only as good as what they know, and llm-wiki is what they'll know.
Team Brain ExrensionClaude - #15formTrust, but check
SECOND ROOM
Second Room is a post-meeting review space for people who weren't in the room. AI turns a meeting into decisions, facts and action items, but instead of asking people to trust the summary, Second Room lets them check each claim against the original transcript and evidence. Multiple reviewers can agree, flag uncertainty or disagree, revealing where AI confidence and human judgment diverge. Thread: Trust, but check. We pulled on the question: what does AI mean for my job when I need to make decisions about a meeting I didn't attend? The product makes AI-generated meeting information traceable, challengeable and reviewable by the people who need to act on it.
2nd ROOMClaude