Atlas

Atlas analyzes GitHub users' public repositories and returns deterministic signal-based profiles, enabling search, comparison, and evaluation.

Project Scope

• CLI: analysis, search, query, inspect, dataset management
• Server: HTTP API for search and analysis
• Frontend: lightweight visualization (see web/)
• Pure pipeline: Acquisition → Normalization → Vestiges → Facts → Indicators → Profile → Evaluation → Projection → Consumers

Commands

• atlas <username> — Analyze a single GitHub profile
• atlas search <query> — Discover developers by intent or expression
• atlas query [options] — Advanced signal-threshold query
• atlas inspect <username> — Inspect a profile in a dataset
• atlas build <usernames> — Build a dataset from usernames
• atlas dataset [info|preview|stats] — Show dataset summary
• atlas serve — Start the HTTP API server

Signal Definitions

• Consistency: recent non-fork repositories divided by max(10, total non-fork repositories); measures recent activity across original work.
• Depth: deep non-fork repositories (size >= 50 KB) divided by max(5, total non-fork repositories); measures substantial original projects.
• Ownership: non-fork repositories with size > 0 divided by all repositories with size > 0; ignores trivial size-0 repositories.
• Activity: based on the latest update across all repositories with graded decay:
  • <= 30 days: 1.0
  • <= 90 days: 0.7
  • <= 180 days: 0.4
  • over 180 days: 0.1

Scoring

Signals are clamped to [0,1] and converted to 0-100 integer component scores. Overall score uses:

• Consistency: 40%
• Ownership: 30%
• Depth: 30%

For very small datasets (fewer than 3 repos), overall score is multiplied by 0.7.

Within a dataset of 10 or more profiles, the final search score is percentile-calibrated; smaller datasets use the raw weighted score.

Architecture

GitHub
   ↓
Transport          internal/github
   ↓
Acquisition        internal/acquisition
   ↓
Normalization      internal/acquisition/normalize.go
   ↓
Vestiges           signals.Repo · profile.UserMetadata · contributions.Summary
   ↓
Facts              signals.Facts
   ↓
Indicators         signals.Signals → RawScore
   ↓
Profile            index.Profile
   ↓
Evaluation         internal/evaluation (overall score, penalty, confidence, ranking policy)
   ↓
Intelligence       Profile + Evaluation output
   ↓
Projection         internal/projection
   ↓
Consumers          cmd/atlas · cmd/server · web

Search execution (internal/engine) sits between Evaluation and Projection, filtering and ranking profiles using the evaluation policy.

Each layer has strict ownership. See docs/ARCHITECTURE.md for layer ownership and docs/INTELLIGENCE.md for the canonical intelligence model.

Search Modes

Use --live to fetch candidates directly from GitHub:

atlas search backend --live

Limited to ~20 users, subject to GitHub API rate limits.

Limitations

• Uses only public repositories.
• No authentication, so GitHub API rate limits apply.
• Repository size is a coarse proxy for project depth.
• Results depend on current GitHub metadata and time window.