Seed Agent

Demo

About

Seed Agent extends Seed Club Talent, an existing community directory, with an evidence-aware talent discovery workflow. Users describe who they need in natural language, find candidates from actual member records, resolve ENSv2 names to wallet addresses, inspect onchain activity through The Graph, and export a candidate report with traceable sources. Profile claims and onchain activity remain separate. AI interprets the search request; it does not invent candidates or missing evidence. Activity scores describe the returned protocol data, not a person's ability, credibility or identity. The current activity query covers Uniswap V3 on Ethereum mainnet. A successful empty query, an unqueried wallet and a failed request are treated as different states. Our demonstration resolves seedclub-ray.eth on ENSv2 Sepolia, matches its address to a member profile, checks protocol activity and exports the results. Address matching is not presented as proof of wallet ownership. The demo wallet has no evidence within the queried protocol scope, which does not imply that it has no activity elsewhere. This is a Continuity project. Accounts, member profiles and basic AI search existed before the event. During the hackathon, we added live onchain evidence, evidence-aware matching, ENSv2 discovery, and candidate reports with inspectable evidence. The repository documents the pre-existing baseline and event development history.

How it's made

The frontend uses Vite and vanilla JavaScript. The backend uses Tencent CloudBase authentication, PostgreSQL and cloud functions, with DeepSeek supporting natural-language processing. The Graph provides indexed Uniswap V3 data from Ethereum mainnet, including recent mint, burn and collect events and position data. Recent events cover a 90-day window; positions and earliest history are queried separately. Deterministic scoring keeps profile relevance separate from activity evidence. For AI searches requiring onchain evidence, the combined match score weights profile relevance at 70% and activity at 30%. We use viem to read ENSv2 through the Universal Resolver on Sepolia. Resolved addresses are compared with member records, and the interface exposes query blocks, resolver information and public name records. Our live demo uses seedclub-ray.eth, which I registered and configured during the event. A page-local shortlist connects discovery, activity checking and Markdown report export. Evidence details include query scope, timestamps, indexed blocks, score components and sample limits. Explorer links are included when transaction evidence is returned. Reports exclude contact fields, and failed requests are never treated as successful empty results. OpenAI Codex generated and revised most implementation code, tests and supporting documentation under my direction. I defined the product goals, made product decisions, performed manual deployments and acceptance testing, registered and configured the ENSv2 name, and recorded the demo narration. The repository documents this substantial AI assistance and distinguishes pre-event functionality from hackathon work.