Vision: Build an AI agent that organizes real-world meetups at ETHGlobal Cannes.
What if AI did not isolate people, but actually coordinated real human interaction?
Most people come to ETHGlobal to learn and meet people, not only to win prizes.
- Everyone wants to connect.
- Nobody coordinates efficiently.
An AI agent that:
- Chats with you (initially via Telegram; Discord is in scope for the product story).
- Understands what you want to learn and what you could teach.
- Matches you with relevant people.
- Organizes real meetups during the hackathon.
Think: AI + messaging + real-world coordination — turning the hackathon into a self-organizing learning network.
| Phase | Goal |
|---|---|
| Friday | Ship MVP: chat agent, matching, meetup proposals |
| Saturday | Onboard as many hackers as possible IRL |
| Later | Layer in World (human verification) and 0G Labs (agent infra) |
This repo starts from the Amazon shopping assistant template used in the AI engineering bootcamp. We adapt it as follows:
- Replace the front-end UI with a Telegram bot as the primary interface.
- Replace the agentic logic with a chat agent that:
- Accepts user requests in natural language, e.g. “I want to learn X and I could teach Y”.
- Appends each request to a global context shared by the agent (later this becomes a Qdrant-backed store for clustering and retrieval).
- Stays polite and does not reveal other users’ data to any single user.
- Uses the updated context to propose learning sessions (e.g. a meetup where one person teaches and a group learns a subject).
- Notifies all relevant participants when a new meetup is proposed.
User identity in context: Telegram handles are stored next to each user’s learn/teach request in the global context so the agent can reach out to the right people when coordinating meetups.
The backend graph is coordinator-led: one turn can chain specialists before the user gets a final answer.
flowchart TB
START([START]) --> COORD[coordinator_agent]
COORD -->|direct reply / plan complete| END([END])
COORD -->|delegate| PI[profile_intake_agent]
COORD -->|delegate| MC[meetup_coordination_agent]
PI --> PITOOLS[profile_intake ToolNode]
PITOOLS --> PI
PI -->|iteration / final_answer| COORD
MC --> MCTOOLS[meetup_coordination ToolNode]
MCTOOLS --> MC
MC -->|iteration / final_answer| COORD
| Node | Role |
|---|---|
| coordinator_agent | Routes the message: answer small-talk directly, or delegate to profile intake and/or meetup coordination (multi-step plans). |
| profile_intake_agent | Parses learn/teach intent and calls append_my_learning_profile to upsert the user in the in-memory global registry (MVP; replace with Qdrant later). |
| meetup_coordination_agent | Calls get_meetup_community_registry to read the full registry for planning only, then create_meetup_proposal_and_notify to record a session proposal and enqueue per-participant notifications (handles included for the Telegram layer). Prompts require privacy-safe replies: no other users’ handles or profiles in the user-visible text. |
Tool nodes mirror the shopping template: each specialist can loop on tools until final_answer is set, then control returns to the coordinator.
Legacy Amazon shopping agent functions remain in agents.py for reference; the live graph no longer uses Qdrant for the final response (that block is commented in graph.py).
API: POST /rag accepts optional telegram_handle on the request body; if omitted, the backend uses thread_id as the display handle.
ETHGlobal Cannes — AI that connects people in the real world.