Talent Discovery

The best engineers aren't looking for jobs. We find them anyway.

Talent Discovery crawls GitHub's social graph starting from 27 known-exceptional seed developers, profiles 3,263 people in their extended networks, scores them with a proprietary "cracked score" algorithm, and surfaces undiscovered talent through an AI-powered chat interface and interactive network visualization.

The Problem

Traditional recruiting finds people who are actively job-seeking. But the best engineers are heads-down building - they have 200 followers, no LinkedIn, and a compiler they wrote from scratch sitting at 15 stars. They're invisible to conventional search.

Our Approach

We flip recruiting on its head: instead of searching by resume keywords, we crawl the social graphs of people we already know are exceptional and use graph analysis + heuristic scoring to find the diamonds in the rough.

Pipeline

27 seed developers (known-exceptional)
        |
        v
  GitHub GraphQL API crawl
  (followers, following, repos, commits, PRs, orgs)
        |
        v
  3,263 profiled developers
  1,874 follow-relationships mapped
        |
        v
  Multi-signal scoring
  - Cracked Score (undiscovered talent metric)
  - Diamond Score (technical depth + low visibility)
  - PageRank & betweenness centrality
  - Louvain community detection
        |
        v
  Two interfaces:
  1. AI Chat Terminal (Claude-powered natural language search)
  2. Force-directed Network Explorer (interactive graph viz)

The Cracked Score

Our proprietary scoring algorithm specifically optimizes for undiscovered talent - not who has the most stars:

• Log-scaled stars/year - a 19-year-old with 500 stars in 2 years scores higher than a senior dev with 10K stars over 15 years
• Youth multiplier - younger accounts with high output get exponential bonuses (accounts <2y get 5x, >12y get 0.3x)
• Famous penalty - >5K followers means you're already discovered (0.7x), >20K gets 0.2x
• Follow-farm detection - high followers with low actual code output gets penalized
• Network bonus - appearing in multiple seed networks (1.5x) or mutual follows with seeds (1.3x) is strong signal

The Diamond Score

A complementary scoring system that finds technical depth others miss:

• Regex-matches repo names/descriptions against 60+ technical signal patterns (compilers, kernels, proof assistants, FPGA, etc.)
• Bonuses for strong languages (Rust, Haskell, OCaml, Zig, Lean)
• HIGH ALPHA flag: technical repos + <200 followers = undervalued
• Academic signals (MIT, Stanford, CMU in bio)
• Cross-network social graph validation

Features

AI Intelligence Terminal

A chat interface backed by Claude that can search, filter, and generate recruiting dossiers in real time. Natural language queries like:

• "Find undiscovered Rust developers with fewer than 500 followers"
• "Who are the youngest developers with the highest cracked scores?"
• "Generate a dossier on @username"

The agent has access to the full dataset and streams responses with tool call visibility.

Network Explorer

Interactive force-directed graph visualization of 400 key developers across the network:

• 400 nodes, 1,874 edges, 4 detected communities
• GitHub avatar rendering with tier-colored rings (seed/tier1/tier2/tier3)
• PageRank and betweenness centrality computed per-node
• Louvain community detection for cluster identification
• Tunable physics (repulsion, link distance, gravity) and display settings
• Filter by tier, language, followers, diamond score, community
• Click any node for full profile details (repos, stats, graph metrics, connected seeds)

Architecture

crawl.py          - GitHub GraphQL crawler (single seed)
batch_crawl.py    - Multi-seed batch crawler with deduplication
scrape.py         - Hackathon project enrichment (repo metadata, contributors)
analyze.py        - Diamond-in-the-rough scoring engine
build_graph.py    - NetworkX graph analysis (PageRank, communities)
server.py         - FastAPI backend + Claude Code SDK agent
app.py            - Modal deployment variant
mcp_talent.py     - MCP stdio server for tool-use integration
index.html        - AI chat terminal frontend
viz.html          - Force-directed network explorer

Zero build step. Every Python script uses uv inline script metadata for dependencies. Frontend is two single HTML files loading libraries from CDN.

Running It

# Crawl a seed's network
uv run crawl.py <github_username>

# Batch crawl multiple seeds
uv run batch_crawl.py user1 user2 user3

# Run the diamond analysis
uv run analyze.py

# Build the graph visualization data
uv run build_graph.py

# Start the main platform
uv run server.py
# -> http://localhost:8000

# Serve the network explorer
python3 -m http.server 8001
# -> http://localhost:8001/viz.html

Dataset

Metric	Value
Seed developers	27
Profiled developers	3,263
Hackathon projects	518
Languages tracked	194
Total stars across profiles	4.8M
Graph nodes (explorer)	400
Graph edges	1,874 (507 mutual)
Communities detected	4

Built With

• Claude Code SDK - AI agent backend for natural language search
• FastAPI - API server
• NetworkX - Graph algorithms (PageRank, betweenness, Louvain)
• force-graph - WebGL force-directed graph rendering
• GitHub GraphQL API - Data source
• uv - Python package management
• Modal - Cloud deployment (optional)