Overview
Seshat Documentation
Technical deep-dives into Seshat's intelligence systems — the scoring engine, the self-learning loop, the discovery layers, and the identity model underneath it all. Written for developers and researchers who want to understand the implementation, methodology, and data pipelines behind the scores.
Start here
See the big picture → Architecture
A top-to-bottom tour of the eight layers — from the UI you're looking at down to the edge runtime it lives on. Read that first, then dive into the details below.
Scoring & Intelligence
How a token becomes a single 0–100 number, and the signals that feed it.
Composite Scoring
How four weighted dimensions (fee 0.20 / dev 0.30 / tech 0.30 / novelty 0.20) blend into the 0–100 composite, plus confidence tiers and hidden-gem detection.
Developer Scoring
Scores the real creator (fee recipient, confirmed by the on-chain claimer) — not the deployer. On-chain track record, FrontrunPro smart followers, background pedigree, activity recency.
AI Research Loop
Claude as an autonomous agent: LARP probability, novelty vs prior art, traction and leadership scoring. Structured JSON stored in R2 + D1.
Fee Scanner
Bankr-API primary discovery, label-aware WETH fee reading on Doppler/V4 pools, V1/V2 ABI fallback, contract-safety checks. The core data pipeline.
Learning Loop
The feedback cycle that makes the scores objectively better over time.
Self-Learning Model
Weekly ridge regression (vanilla TypeScript, zero ML deps) recalibrates the four dimension weights from labelled outcomes. Cross-validated R² kept honest.
Outcome Research
The tiered auto-labeler that measures what actually happened to each scored token — rules for the obvious, Claude-with-tools for the ambiguous — feeding the learning loop.
Discovery & Data
Surfacing where attention concentrates, finding by meaning, and cross-source context.
Trending Narratives
Clusters recently-scored tokens into trending market narratives with a momentum score — proactively surfacing where attention is concentrating.
Semantic Search & RAG
Workers AI embeddings (bge-small, 384-dim) + a Vectorize index power “similar tokens” and the retrieval-augmented chat.
Multi-Persona Swarm
Five opinionated Claude personas debate high-conviction tokens into a sentiment read and a 0–100 crowd score — a captured signal, not yet a scoring input.
Market-Data Enrichment
CoinGecko + DeFiLlama add categories, ATH data, an independent confidence-rated price, and protocol TVL as cross-source context.
Identity & Safety
The principle the whole system is built on.
Pipeline at a glance
Seshat runs entirely on Cloudflare Workers (TypeScript/ESM) with zero server infrastructure. All compute is edge-native; persistence is Cloudflare D1 (SQLite) and KV; blob storage is R2; semantic search runs on Workers AI + a Vectorize index. Three cron triggers drive everything, fanning heavy work out through Cloudflare Queues.
Admin endpoints
All require an Authorization: Bearer token.