What shipped, and what's landing.

Share a hosted form link or drop the one-script embed widget into any page.

Respondents record a spoken answer in the browser; it's transcribed, redacted, and extracted into a tidy row.

Describe your intake out loud and watch a real, shareable form assemble in a live panel.

The form answers respondent questions from your uploaded documents, cites the source, and refuses plainly when it doesn't know.

PII is redacted into a keyed vault upstream of the model, and content is moderated both ways.

HMAC-signed webhook delivery, a submissions dashboard, and CSV / NDJSON exports.

A real-time voice agent with barge-in and backchannel acknowledgements. Rolling out on the Business tier.

An edge failover lane and a regional PII sidecar for EU data residency. A locked architecture, hosting in progress.

Describe what you want, see a live preview, refine it in words, and approve — prototyped and rolling out.

Accept voice-note intake over WhatsApp and SMS. Planned for a v1.x release, not built yet.

The engine is multilingual by design; the customer-facing language rollout is planned, with the infrastructure validated dormant.

Hot-swappable premium voices for the answer-back and live agent. Deferred to a later release.

Measuring the self-hosted inference cost per response end to end. Until it's validated, the cost advantage is stated as architecture, never a number.