Conversate

What is live now

Working product: Conversate is a deployed two-person browser voice room with private invitations, adult consent, guided speaker turns, and generated translated audio.

Gemini’s role: An operator-triggered Gemini 3.7 Flash call converts aggregate, redacted operating metadata into a schema-validated draft brief that is saved pending human review.

Business model: Owner-paid at $6.99/month for 60 translated minutes; guests join free; no overage charges.

Evidence baseline: Verified external users: 0; paying users: 0; earned revenue: $0; public testimonials: 0; independent accuracy benchmark: none.

The problem

Conversate began with a family conversation Brian Smith wanted to have more naturally with members of his wife’s Filipino family. Phrase translation helped with words, but a remote call still became fragmented or depended on a bilingual relative.

The commercial hypothesis is a recurring Small Business Services workflow: an owner coordinating with a customer, contractor, or vendor who prefers another language. The target is ordinary, low-stakes conversation—not medical, legal, financial, emergency, or other high-stakes decisions.

The working product

An owner creates a private translated-voice room and sends a one-time invitation. The guest joins free in a browser. Each participant accepts the room notices and sets up audio. Conversate then guides one speaker at a time: the active speaker holds to talk, the service sends that turn to the configured translation route, and the listener receives generated translated audio.

The production customer and judge voice route uses app-owned OpenAI Realtime. Judges receive free, time-bounded access to that same product and route through private submission instructions; it is not a competition-only voice build. A controlled production capture shows the product moving through ready, listening, translating, translated-audio arrival, and return to ready. A separate two-device founder check completed English-to-Filipino and Filipino-to-English turns. That is a bounded founder observation, not an independent accuracy benchmark.

Gemini Live voice was tested separately, sounded garbled to the founder, and was rejected before customer or judge traffic. It is not the current voice route.

Gemini runs the management-analysis loop

Gemini 3.7 Flash is live in the IAP-protected management backend. On an authorized refresh, the application assembles aggregate, redacted operating metadata, excludes judge activity from customer evidence, and calls the Gemini API with a required JSON schema. The application rejects malformed or privacy-regressing output, saves a valid run with source and model provenance, and marks it pending human review.

Gemini performs the bounded analysis task: identifying risks, next actions, finance notes, feedback trends, invite friction, and test hypotheses. It cannot spend, change provider routing, contact customers, approve a claim, publish output, accept terms, or submit the entry.

AI-native operations, bounded by human authority

AI is embedded in two operating layers. In the product, a voice model executes translated-audio turns within application-controlled identity, consent, speaker, route, model, prompt, fallback, usage, and retention boundaries. In management, Gemini performs schema-bounded operating analysis. AI-assisted engineering and QA workflows also help the sole founder research, implement, test, inspect, and reconcile evidence. Brian retains authority over credentials, spend, production changes, pricing, refunds, privacy, legal and eligibility positions, customer evidence, public claims, and publication.

Conversate does not retain raw microphone audio or transcripts in its application by default. Raw microphone audio is not forwarded directly to the other participant. “Private” describes invitation and access design; it does not mean end-to-end encryption.

Conversate has no employees or contractors. If customer evidence later supports growth, the business could create paid work in language-quality evaluation, accessibility testing, customer support, community distribution, and workflow integration. Those are potential roles, not current jobs or partnerships.

Business model and evidence baseline

The working model is owner-paid and guest-free: $6.99 per month for 60 translated minutes, no guest subscription, and no overage charges. Retention depends on a recurring relationship or business job, reliable two-way completion, transparent limits, trust, and support burden.

As of August 15, verified external users, paying users, earned revenue, and public testimonials were all zero. A $6.99 founder technical payment was fully refunded and is not revenue. Verified project-attributed incurred Google Cloud expense was $130.85; with no external service delivered, it is classified as R&D/hosting rather than customer COGS. No paid acquisition campaign launched, so CAC is undefined and demand remains unproven.

Next falsifiable business test

The next gate is one consenting arms-length small-business workflow and one consenting cross-cultural family pair. Conversate will measure invite completion, successful translated turns in both directions, support effort, participant-rated meaning, tone, and trust, voluntary full-price payment, repeat use, and a specific reason for non-repeat. Paid acquisition starts only after an organic segment and message activate users, with a capped budget and predeclared CAC stop rule.

Pre-window family-translation prototypes and later prompt-text reuse are fully disclosed in Additional Information. The required disclosure is included; final eligibility remains with the Sponsor and Administrator.

Conversate is not claiming adoption or product-market fit. It is claiming a working product, a live Gemini management-analysis loop, exact evidence boundaries, and a disciplined next test.

Built With

• artifact-registry
• cloud-build
• cloud-run
• cloud-sql
• fastify
• gemini-3-7-flash
• gemini-api
• openai-realtime
• postgresql
• react
• secret-manager
• stripe
• typescript