You can’t “AI strategy” your way to product-market fit. You have to measure your way there.

Just dropped a new episode of Skywatch (Bluejay (YC X25)'s car podcast on conversational AI) with Kylan Gibbs, CEO & co-founder of Inworld AI — and we went deep on something I think the whole industry is about to relearn:

Most teams are still shipping voice + agents based on vibes (“this sounds good to me”)…

…but their users don’t think it does.

This convo felt extra timely because Inworld shipped TTS-1.5 last week — a very “benchmarks-first” release: improved expressiveness + lower word error rates, plus real-time latency/cost targets that actually make consumer-scale voice feasible.

A few takeaways I’m still thinking about:

🔹 Quality is not universal. “Good” depends on the customer, the demographic, and the context (support vs health coach vs gaming).
🔹 Offline benchmarks help, but online experiments are where the truth comes out. A/B tests + real user outcomes > opinions.
🔹 The pipeline matters more than any single model. STT ↔ LLM ↔ TTS are coupled — the “best” component alone rarely wins.
🔹 Cost + latency unlock new products. When voice gets cheap + fast enough to iterate, experimentation stops being a luxury.

If you’re building in voice, real-time infra, or agentic systems, you’ll love this one.