Hackathon Portal
AI Tinkerers - Prague
Sign in

Generative UI Global Hackathon: Agentic Interfaces

Sign in
Engineering Lead at Stealth AI Startup - Stealth AI Startup
Own AI-native architecture and backend pipelines remotely with no equity gimmicks.
All Jobs
Team

Timmy´s Clowns

Project Concept

Coder’s Magic Cookie is a fortune-teller of bugs. You describe how you’re stuck — “the test passes locally”, “this fix keeps not fixing it”, “I can not install my usb mouse” — and a Gemini-backed LangGraph agent answers in three short paragraphs:

~70 words of brutally funny slapstick about a fictional dev who suffered worse (Mira of the third-floor coworking, punching the monitor and bruising her fist; Joren of the converted closet, headbutting the keyboard until it printed fifteen pages of aaaaa).
A bold cryptic aphorism in broken-English fortune-cookie cadence, signed by an invented master who never repeats — “Master Vex of the Cracked Compiler say: When bug hide, lamp also hide.”
A real, concrete technical fix for the user’s actual situation — woven into italic prose, no bullet lists, no code blocks. The user can extract a real action from it.
The premise is that humor and a soft re-frame restore the momentum debugging steals. You walk away with both a smile and a lead.

The agentic-interface angle

We took the CopilotKit hackathon starter kit and added a second LangGraph graph (graphId: “story”) sharing the same Gemini-Flash runtime as the kit’s lead-triage agent — but with no backend tools and a swapped system prompt. The BFF registers both agents under one CopilotKit endpoint. A new /story route uses to drive it. About 80 lines of net new code: one Python prompt module, one BFF agent registration, one Next.js route. The whole thing demonstrates how cleanly the kit's AG-UI rails support adding a sibling agent without touching the existing canvas.

Tech stack

CopilotKit v2 (chat + AG-UI runtime) · LangGraph (two graphs in one process) · Google Gemini 3.1 Flash-Lite · Hono BFF · Next.js 15 / Turbopack · Notion MCP (existing kit feature, unchanged) · Manufact / mcp-use (existing) · Postgres-backed thread persistence

Repo

https://github.com/jakubbinter-gc/Generative-UI-Global-Hackathon-Starter-Kit/tree/claude/funny-spence-694fc3

Two commits on top of main:

fix: Windows-cross-platform dev scripts + langgraph blocking flag — three port/script defaults that block a fresh clone on Windows
feat: Coder’s Magic Cookie — story-doctor agent at /story — the feature itself

Entry

Status: Submitted

Last saved: May 09 at 7:05 PM CEST

View Entry

Team Roster

You must be registered for the event to view the team message board.

Jakub Binter Team Lead RSVP Approved

Researcher, Analyst at Vltava Labe Media, Univerzita Karlova, Univerzita Jana Evangelisty Purkyně
Solo team — one human, many cookies. Jakub set the product concept (Coder's Magic Cookie: a fortune-teller-of-bugs that turns developer despair into a 130-word slapstick story plus a real fix), wrote the system prompt that gives the cookie its voice, and shipped end-to-end on top of the Generative UI Global Hackathon Starter Kit. No actual cookies were harmed during development. Several real ones were eaten. CopilotKit (v2) — added a new /story route using <CopilotChat agentId="story"> with broken-English suggestion chips ("Bug no die", "Cursor blink at me", "Refactor eat me"). Also debugged three out-of-the-box port/script defaults that block the kit from booting on Windows. Master Phin of the Empty Pipe approved. LangChain / LangGraph — registered a second graph (story_graph) in langgraph.json and apps/agent/main.py sharing the same runtime as the kit's lead-triage agent but with a different system prompt and zero backend tools. Wired as a second LangGraphAgent in the Hono BFF. The kit now hosts both a serious lead-triage assistant and a small wise cookie that talks to it from the next desk over. Google DeepMind / Gemini — Gemini 3.1 Flash-Lite is the LLM behind both agents. Tuned the system prompt across many takes until Gemini stopped writing 600-word three-act dramas and started writing 130-word fortune cookies. Took longer than admitting one's USB mouse is, in fact, plugged in. Manufact / mcp-use — kept and exercised the kit's existing Notion MCP integration during the demo (the Story-Doctor uses no MCP tools; the lead-triage canvas pulls 50 real workshop signups from Notion to prove the rails work). Coding was AI-assisted via Claude Code; design, prompt-engineering, cookie-tasting, and shipping decisions were the human's. Total feature delta: ~80 net new lines + ~20 lines of kit fixes, in three commits on a forked branch. Master Vex of the Cracked Compiler say: "Code that ship is code. Code that not ship is also code, but sad."
Jakub Binter is a behavioral scientist and researcher associated with the Faculty of Social and Economic Studies at UJEP and the Faculty of Science at Charles University. His research focuses on human behavior, decision-making mechanisms, and psychophysiological responses to external stimuli. He is employing experimental psychology methods to assess cognitive and affective processing. His work frequently integrates AI-driven analytics and statistical modeling to refine the understanding of human perception by combining evolutionary psychology with neuroscience-based approaches. Currently he works as content analyst for Vltava Labe Media.
AI, Generative AI, Machine Learning, LLMs, Agentic LLMs, RAG, Databases, SQL, Model Training
I work in multiple research fields on two universities with international collaboration where I lead a small research team. Our main focus is perception research, automation, human computer interaction.