# EdgeGate

> Hardware-in-the-loop CI/CD platform for edge AI. Automated regression testing on real Snapdragon devices through Qualcomm AI Hub.

EdgeGate is a testing and CI/CD platform that validates AI model performance on real Qualcomm Snapdragon hardware before deployment. It catches regressions that cloud benchmarks miss — thermal throttling, firmware quirks, quantization drift, and NPU fallback issues.

## Key Capabilities

- **Real Device Testing**: Models run on physical Snapdragon chipsets (8 Gen 3, 7s Gen 2, etc.) via Qualcomm AI Hub, not emulators
- **Automated Quality Gates**: Define pass/fail thresholds for inference time, memory usage, and NPU utilization. Gates block PRs on regression
- **Signed Evidence Bundles**: Every test run produces Ed25519-signed, SHA-256-hashed evidence reports for auditability and compliance
- **CI/CD Integration**: GitHub Actions integration with HMAC-SHA256 authenticated API. Results appear as PR checks
- **Deterministic Testing**: Median-of-N gating, warmup exclusion, and flake detection bring statistical rigor to non-deterministic inference
- **Multi-Tenant Workspaces**: RBAC, API keys, and workspace isolation for team collaboration

## Pricing

- **Playground** (Free): 10 runs/month, 1 workspace, 2 devices per run
- **Pro** ($49/month): 100 runs/month, 3 workspaces, GitHub Action, flake detection, signed evidence bundles
- **Team** ($149/month): 500 runs/month, 10 workspaces, RBAC, API access, webhooks, audit logs

## Documentation

- [CI/CD Integration Guide](https://edgegate.ai/docs/integration): Step-by-step GitHub Actions setup with API reference
- [Benchmarks](https://edgegate.ai/benchmarks): Sample benchmark report showing FP32 vs INT8 person detection on Snapdragon 8 Gen 3

## Blog

- [The Hidden Cost of Edge AI Regressions](https://edgegate.ai/blog/the-hidden-cost-of-edge-ai-regressions): Why INT8 models break on real devices and how quality gates prevent production failures
- [Deterministic Testing for Non-Deterministic Models](https://edgegate.ai/blog/deterministic-testing-for-non-deterministic-models): How median-of-N gating and flake detection bring statistical rigor to hardware testing
- [Evidence Bundles: Software Release Rigor for ML](https://edgegate.ai/blog/evidence-bundles-software-release-rigor-for-ml): Cryptographically signed proof that models passed quality gates on real hardware
- [Why Cloud Benchmarks Lie About Edge Performance](https://edgegate.ai/blog/why-cloud-benchmarks-lie-about-edge-performance): The gap between cloud inference times and real Snapdragon performance
- [Hardware-in-the-Loop Testing for AI](https://edgegate.ai/blog/hardware-in-the-loop-testing-for-ai): Why emulators miss thermal throttling, firmware quirks, and quantization drift
- [Building a CI/CD Pipeline for On-Device AI](https://edgegate.ai/blog/ci-cd-pipeline-for-on-device-ai): Step-by-step guide to regression gates on real Snapdragon hardware in every PR

## Contact

- Website: https://edgegate.ai
- Email: hello@edgegate.dev