Why we shipped Walkie-Talkie before Live Conversational

Context

Celer has two audio-practice modes on the roadmap:

• Walkie-Talkie — turn-based. You speak, the question stops, the model thinks, the answer plays. Like a slow phone call.
• Live Conversational — full duplex. Streaming STT in, streaming TTS out, interruptible, with latency under a couple of hundred milliseconds. Like a real interview.

Both are good features. Live Conversational is the one that gets demoed.

Decision

Ship Walkie-Talkie first. Treat Live Conversational as a separate, later product surface.

Consequences

Walkie-Talkie ships in roughly a week. No realtime infrastructure. Standard HTTP. STT and TTS as discrete calls. Bandwidth is a non-issue, jitter doesn't exist, and there's nothing to debug at 3am.

Live Conversational stays scoped. A separate, premium tier when it does land. Premium pricing is much easier to defend on a feature that needs WebRTC, streaming inference, and a media bridge running 24/7.

Users get a useful audio mode now. Most candidates practising for an interview will get 80% of the benefit from turn-based. The interview itself isn't a duplex barrage; it's a slow rhythm of question → think → answer.

What this costs

Four weeks of "real" audio mode is delayed. For some users, that's the killer feature. They'll have to wait — and we say so on the roadmap.

Why it was tempting to do the wrong thing

Live Conversational is what the demo looks impressive on. There is a real, sticky pull toward building the demo-friendly thing first. Acknowledging that pull, and shipping the cheaper-but-useful thing anyway, is most of the ADR.

Alternatives considered

• Both at once. Two modes plus the rest of the product is too much for a single operator in any reasonable window.
• Live Conversational first. Ego-driven roadmap. Killed.
• Walkie-Talkie only, ever. Cost: leaves premium tier without a clear differentiator. Killed.