Architecture

Overview

Ingestion Engine: Pulls live signals from Twitter, Farcaster, Google, and new token listings (multi-chain). Applies normalization, deduplicates, and writes to Redis streams and PostgreSQL
Scoring Engine: Processes raw events and normalizes into canonical topics. Calculates scores using a sophisticated weighting strategy
Agent: Reads processed trend data and decides what should go onchain. Powered by modular tools that expose underlying data access - DB queries, API calls, Clanker SDK, image generation, etc.

Sources: Twitter, Farcaster, Google, new token listings from major chains. Starship tracks both accounts of interest and global trends. Soon: Reddit, 4chan, News.
Quality Controls:
- Recency filters
- Deterministic deduplication across sources
- Topic extraction that handles recasts, quotes, images/videos.
Output:
- Redis streams for the real‑time pipeline
- PostgreSQL as the durable event log (and everything else).

Normalization & Mapping: Events are normalized, then resolved to canonical topics
General Scoring: Recency‑weighted engagement with additional source/author weights
Token Scoring: Dedicated path for token events:
- Multi‑chain awareness (e.g., Base, Solana, Ethereum) and topic similarity aggregation
- Quality signals and chain weights
- Used as a popularity signal for matching emerging trends.

Inputs: Multi-layer - top-level trends, followed by tool-usage for context building based on raw events, social media search, etc.
Decisions: Focuses on evidence at hand, mixed with some baked-in degeneracy for a sprinkle of unpredictability
Outputs: Fully automated deploys, with artifacts pinned to decentralized storage
- Generates token metadata
- Clanks the token on Base
- Posts to Farcaster. Soon: Twitter
Social Interactions: The agent reads all mentions (Farcaster-only on launch). It mainly cares about current affair chat (crypto and global) and replies at its own discretion.