TalkaFlow: scaling a multi-tenant voice-AI platform

TalkaFlow is a multi-tenant voice-AI platform built on a single core that serves five quite different product lines: clinic phone agents with practice-management integration, telephony, booking and recall; companion devices with a persona and persistent per-device memory; a self-service check-in kiosk with insurance verification; a reception-desk screen pop; and an emerging live interpreting product.

Three environments, one codebase

The platform runs in production in Sydney, on a second fully independent production stack serving another region, and on a staging box used to rehearse every change before it ships. Billing is live, managed model-key pooling runs on both production stacks, and the interface is multilingual from the outset.

From feature-rich to production-grade

Growth to more clinics and a fleet of more than a thousand companion devices raises the bar on the parts of a platform users never see. Ahead of that scale-up the system is going through a formal production-readiness review — the same discipline we would apply to any platform carrying clinical workloads: how secrets are held, how data is backed up and restored, how releases roll forward and back, and how failures are detected and escalated. The findings are worked as a prioritised engineering queue rather than a retrospective.

And the product backlog

• Metered billing for the check-in kiosk and per-organisation product entitlements.
• Companion-device work: emotional expression, animated persona, tool use, retrieval-grounded responses and wake-word.
• Clinic recall, plus payment-terminal certification against a fixed external deadline.
• Voice and model selection surfaced in the interface instead of in configuration.
• A Raspberry Pi edge gateway, and productising the live interpreter.

Ways of working, made explicit

The last workstream is the one that makes the others repeatable: staging first, pull-request review, internationalisation from the outset, encrypted secrets and documentation rules — written down and enforced, so four engineers can work in parallel on the same core without regressions finding their way to production.