VivaGuard · Deployment Options
VivaGuard — Choosing How to Run It
Three ways to deploy VivaGuard voice-integrity for online exams & interviews.
| Deployment option | Tech spec1 | Where your data lives | Who runs it | Time to launch | Scaling | Available today |
|---|---|---|---|---|---|---|
| On-premise in your own data centre |
1 server (4–8 vCPU, 8–16 GB) for a pilot; add a GPU box (1–2× NVIDIA L4) for exam-mode at full scale. Docker-based; can run fully offline / air-gapped. | Entirely inside your network — nothing ever leaves your servers. Simplest path for DPDP / biometric compliance. | You provide and run the hardware, updates and backups. | Longest — hardware, install, and your security review. | Limited by the hardware you provision. | ✅ Ready to install now |
| In your own Google Cloud | Runs in your GCP project (Cloud Run or GKE): a 4 vCPU / 8 GB API tier plus a 2–4× NVIDIA L4 GPU pool for exam-mode; voiceprints in your Cloud SQL or Cloud Storage. | Stays in your own cloud project and region — you keep full control and data residency. | Your cloud team runs it; we help you deploy and can co-manage. | Moderate — a cloud deployment into your project. | Elastic — scales up and down automatically. | ✅ Ready (with a small storage add-on for scale) |
| Fully managed by Dheemai | Runs on Dheemai's autoscaling GKE + GPU pool; you simply call the API with your key. | Candidate audio is processed on our infrastructure — needs a data-processing agreement and explicit consent. | We run everything — scaling, updates, monitoring. You manage nothing. | Fastest — live as soon as we issue your API key. | Elastic — we scale it for you. | 🔜 On our roadmap (multi-tenant service in development) |
Footnotes & assumptions
- Tech spec is sized for up to ~1,000 concurrent exam sessions at a peak of roughly 10–30 exam-mode analyses per second. A small pilot (a few dozen users) runs comfortably on a single CPU server with no GPU.
- Exam-mode (analysing a spoken answer for reading vs. genuine speech) is the compute-heavy part — ~60 seconds of CPU per 60-second clip, or ~1–2 seconds on a GPU. Identity verification is light (~1–2 seconds) and needs no GPU.
- GPU figures are based on NVIDIA L4-class accelerators; counts scale roughly linearly with peak exam-analyses-per-second.
- “Available today” reflects the current single-tenant product. The fully-managed, multi-tenant service requires additional engineering before launch.
- DPDP = India's Digital Personal Data Protection Act; a person's voice is treated as sensitive / biometric personal data, which is why data location matters.
Our recommendation: for production with real candidate voices, choose On-premise or In your own Google Cloud — your data stays under your control and both are available today. Use the fully-managed option for a quick, consented pilot while you evaluate.