Independent automated benchmarking of frontier AI models. Tracks performance regressions so the community knows when models degrade.
- Website: Next.js 15 (App Router), TailwindCSS, Recharts, Framer Motion
- Benchmark Engine: Python 3.13, SQLite, runs on DGX Sparks
- Deployment: Linode server, PM2, Nginx reverse proxy
npm run dev # Dev server on port 3002
npm run build # Production build
npm start # Production server on port 3002
bash deploy.sh # Deploy to productioncd benchmark
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python runner.py --schedule morning # Run benchmarks
python export_json.py --output ../public/data # Export JSONSingle daily run at 3am ET via benchmark/run_benchmarks.sh. Outage monitor every 30 min. PATH must include ~/.local/bin and ~/.npm-global/bin for CLI tools.
app/— Next.js pages (dashboard, models, categories, compare, evidence, outages, methodology, about)components/— React components (shared, charts, dashboard)lib/— Types, utilities, data loading, model/category metadatapublic/data/— JSON data files (generated by benchmark engine)benchmark/— Python benchmark suite (config, db, runner, tests, export)config/— Nginx configuration templatedeploy.sh— Blue-green atomic deployment script
Claude Opus 5, Claude Sonnet 5, GPT-5.6 Sol, Grok 4.5. Claude aliases and Codex/Grok catalogs resolve at runtime so scheduled benchmarks follow each provider's current frontier model.
benchmark/runner.py— Main benchmark orchestratorbenchmark/export_json.py— SQLite to JSON exporterbenchmark/run_benchmarks.sh— Cron entry pointdeploy.sh— Production deploymentecosystem.config.js— PM2 process config