GitSyntropy is a multi-agent system that scores team compatibility and simulates hiring impact using GitHub behavioral data and psychometric profiling. It surfaces concrete risks and recommendations, not just raw metrics and numbers.
Engineering teams fail due to misalignment in work patterns, communication styles, and decision-making : not lack of technical skill. Existing tools like MBTI and DISC rely on self-report surveys that are subjective, static, and disconnected from real work behavior.
GitSyntropy derives compatibility signals from what engineers actually do: commit timing, PR activity, collaboration frequency, and self-assessed behavioral dimensions. It produces a continuously-updatable model of how a team is likely to function together.
Team health check Input GitHub usernames. Get a compatibility score, dimension-level breakdown, and flagged risks.
Hiring decision support Simulate adding a candidate to the current team. See whether the score improves and which friction points get resolved.
Async optimization Detect chronotype gaps between team members. Recommend working models based on peak-hour overlap and collaboration index.
Team: atlas-core (4 members)
Compatibility Score: 24.3 / 36 (Good)
Dimension Breakdown:
Chronotype Sync 6 / 8 WARNING -- 6h peak-hour gap
Stress Response 6 / 7 OK
Risk Tolerance 4 / 6 OK
Decision Style 3 / 5 WARNING -- high variance
Work Style 4 / 4 OK
Team Resilience 1 / 3 LOW
Leadership Spread 1 / 2 OK
Innovation Drive 0 / 1 OK
Risks:
- Chronotype mismatch: synchronous standups will be low-quality
- Decision style variance: likely friction under deadline pressure
Recommendations:
- Shift to async-first workflow
- Target hire profile: early bird, low risk tolerance, structured thinker
Hire simulation
Current score: 24.3 / 36
With candidate X: 29.1 / 36 (+20%)
Candidate resolves: Chronotype mismatch, resilience gap
Candidate introduces: Minor leadership overlap (acceptable)
GitHub API
|
v
GitHub Analyst Agent -- commit patterns, PR activity, chronotype extraction
|
v
Psychometric Profiler -- 8-dimension behavioral profile from adaptive assessment
|
v
Compatibility Engine -- variance-based weighted scoring across 8 dimensions
|
v
Monte Carlo Simulator -- candidate hire impact simulation (N iterations)
|
v
Synthesis Agent (Claude) -- narrative report: risks, recommendations, hiring gaps
|
v
WebSocket stream -> Frontend (Astro + React)
Backend: FastAPI (async), SQLAlchemy + asyncpg, Supabase (PostgreSQL), LangGraph orchestration Frontend: Astro, React, Tailwind CSS, Framer Motion, recharts AI: Anthropic Claude (streaming synthesis), LangGraph multi-agent pipeline Infra: Railway (backend), Vercel (frontend), GitHub Actions CI/CD
ML and statistical modeling
- Variance-based weighted scoring model across 8 behavioral dimensions
- Monte Carlo simulation for hire impact prediction (configurable iteration count)
- Behavioral feature engineering from raw GitHub telemetry
Data engineering
- GitHub API ingestion pipeline with rate-limit handling
- Feature extraction: commit timestamp distributions, PR merge patterns, collaboration index
- Chronotype derivation from commit activity heatmaps
System design
- Multi-agent orchestration via LangGraph with step-level streaming
- WebSocket-based real-time pipeline progress to the frontend
- Computerized Adaptive Testing (CAT) algorithm for psychometric assessment
Backend engineering
- FastAPI async API with session-mode PostgreSQL (not pooler) via asyncpg
- GitHub OAuth 2.0 flow with JWT session management
- SQLAlchemy async ORM, Alembic-ready schema
Frontend engineering
- Astro islands architecture with selective React hydration
- nanostores for cross-island shared state
- Framer Motion page and card animations with AnimatePresence
The engine scores 8 behavioral dimensions derived from the assessment and GitHub signals. Each dimension carries a weight (1-8) calibrated so that high-stakes alignment factors (chronotype, stress response) contribute more to the total.
Scores reflect not just averages but variance, two team members with opposite profiles on a high-weight dimension trigger a risk flag even if the mean score is acceptable. This variance sensitivity is what separates the model from a naive sum.
Maximum possible score: 36 points.
cd apps/backend
python -m venv .venv
.venv\Scripts\activate # Windows
pip install -e ".[dev]"
cp .env.example .env # fill in Supabase URL, Anthropic key, GitHub OAuth
uvicorn app.main:app --reload --port 8000Required .env values:
GS_DATABASE_URL=postgresql+asyncpg://... # Supabase direct connection, port 5432
GS_ANTHROPIC_API_KEY=sk-ant-...
GS_GITHUB_CLIENT_ID=...
GS_GITHUB_CLIENT_SECRET=...
GS_JWT_SECRET=... # any strong random string
cd apps/frontend
npm install
npm run dev # starts on http://localhost:4321Frontend reads PUBLIC_API_URL from .env (defaults to http://localhost:8000).
GitSyntropy/
apps/
backend/
app/
main.py -- FastAPI routes
services.py -- business logic, agent calls, scoring
database.py -- async engine, session factory
models.py -- SQLAlchemy ORM models
schemas.py -- Pydantic request/response schemas
config.py -- env-based settings (pydantic-settings)
frontend/
src/
components/ -- React island components (Dashboard, Assessment, etc.)
lib/ -- API client, nanostores, motion variants
pages/ -- Astro pages (SSR shell + React hydration)
.github/
workflows/
deploy.yml -- Railway (backend) + Vercel (frontend) deployment
Backend deploys to Railway, frontend to Vercel. Both trigger on push to main.
GitHub Actions secrets required:
| Secret | Where to get it |
|---|---|
RAILWAY_TOKEN |
Railway dashboard -- Account -- Tokens |
PUBLIC_API_URL_PROD |
Your Railway service URL |
BACKEND_URL |
Same as above (used for health checks) |
VERCEL_TOKEN |
Vercel dashboard -- Account -- Tokens |
VERCEL_ORG_ID |
Vercel project settings |
VERCEL_PROJECT_ID |
Vercel project settings |
Backend environment variables (set in Railway dashboard, not GitHub secrets):
GS_DATABASE_URL, GS_ANTHROPIC_API_KEY, GS_GITHUB_CLIENT_ID, GS_GITHUB_CLIENT_SECRET, GS_JWT_SECRET
Core pipeline functional: GitHub sync, adaptive assessment, compatibility scoring, Monte Carlo simulation, Claude synthesis, WebSocket streaming.
In progress: Hire simulation UI on the compatibility page.