Or jump from this list on a smaller screen.
This is a page from the Audivra documentation. Developer exports — index: http://127.0.0.1:3000/docs/llms.txt, full corpus: http://127.0.0.1:3000/docs/llms-full.txt, OpenAPI: http://127.0.0.1:3000/docs/openapi.json, Postman: http://127.0.0.1:3000/docs/postman.json. Press ⌘K to search docs.
Enterprise & private deployment
Air-gapped inference, on-prem GPU clusters, SLA-backed uptime, and VPC isolation — for teams that need voice + avatar APIs inside their own perimeter.
Why private deployment
Audivra's self-hosted stack is stronger than a generic "private deployment" blurb — fast-path routing, unified voice + avatar API, and honest processing-node docs. Enterprise customers extend that with SLA-backed uptime, VPC isolation, and air-gapped inference for regulated industries.
SLA tiers
| Tier | Uptime | Support | Notes |
|---|---|---|---|
| Standard (cloud SaaS) | 99.5% monthly | Email · docs · community | Shared multi-tenant — audivra.dev or your custom domain |
| Enterprise (private cloud) | 99.9% monthly | Dedicated channel · onboarding · quarterly review | Single-tenant VPC, optional air-gapped inference tier |
| Air-gapped (on-prem) | Customer-operated | Runbook + remote break-glass (optional) | No outbound internet from inference tier; license key offline validation |
Architecture layers
Private VPC (recommended B2B)
Audivra SaaS + gateway in your AWS/Azure/GCP VPC. Postgres (Supabase self-hosted or RDS), Redis, and S3-compatible storage stay inside the account. Modal/RunPod GPU calls optional or replaced with on-prem nodes.
- Deploy docker-compose.prod or K8s helm chart in private subnet
- Restrict ingress to corporate VPN or Zero Trust proxy
- Use customer-managed KMS for S3 and database encryption
- Route /api/v1/* through internal ALB — no public dashboard unless required
On-prem GPU cluster
Dedicated RTX/A100 nodes run inference/main.py and local_worker.py. Fast-path routing sends short prompts to local Kokoro/Ollama; heavy lip-sync stays on-cluster MuseTalk workers.
- Minimum: 1× RTX 4090 (24 GB) for MuseTalk + Kokoro concurrent load
- Scale: Celery worker pool bound to GPU count — one video job per GPU slot
- Set PREFERRED_NODE=local_worker and disable external_api in routing profile
- Monitor queue depth via Redis + /app/admin (or Prometheus exporter)
Air-gapped / offline
Inference tier has no egress. Model weights and voice assets ship on signed media; billing and admin UI run on a separate connected zone or entirely offline with manual license files.
- Set AIR_GAPPED=true or DEPLOYMENT_MODE=air_gapped in .env
- npm run inference:bootstrap on a connected staging machine, then transfer artifacts
- Disable Modal, RunPod, Stripe, and external TTS fallbacks in env
- Use WORKER_SECRET for worker-to-gateway auth only
- Audit webhook and payment routes — disable if not applicable offline
Security & compliance
| Control | Detail |
|---|---|
| API keys | SHA-256 hashed at rest; rotate via /app/keys or admin API |
| Data residency | Self-host Postgres + S3 in your region; no cross-border inference by default |
| SSO / SAML | Entra ID, Okta, or Google Workspace via Supabase Auth (enterprise config) |
| Admin audit | /app/admin usage console; webhook_events + job_ledger tables for billing disputes |
| Rate limits | 60 req/min per key — tunable per tenant in private deployments |
Verify enterprise install
npm run verify:enterprise npm run verify:production npm run verify:fast-path npm run build
Contact sales
Enterprise onboarding
Include expected monthly characters, video minutes, GPU footprint, and compliance requirements (SOC 2, HIPAA BAA, etc.).
Related: Deploy checklist · RTX 4090 guide · Processing nodes