Hinglish is a smart call, most voice agents fall apart the moment people switch languages mid-sentence. How are you handling code-switching in the ASR, one multilingual model or detecting per turn?

That 40% after-hours number makes sense. The 9pm call is often the urgent one too, like a burst pipe or no heat, so whoever answers first basically wins the job.

Nice work, and keeping it read-only makes it much easier to trust. Are the SAOS court rulings searchable by full text, or mostly by case metadata?

Agreed, and a lot of the long tail doesn't even have a portal. The plumber, the clinic front desk, the local restaurant just have a phone number. Browser agents plus phone calls covers most of what's left.

Congrats on the talk. Spec to server is such a practical on-ramp. Curious how mcpsmith handles big specs: does it expose every endpoint as a tool, or let people pick a subset so the agent isn't swimming in tools?

Great catch, and a sneaky one. A small cleanup pass between the LLM output and TTS pays off for exactly this reason. Ellipses, slashes, and URLs are worth testing too, since models love them and they rarely read well on a call.

Same. Hold time is pure dead time, and it's one of the easiest things to hand off because the goal is so clear. That's a big part of why we built Patch: your agent calls from your own number, works through the phone menu, waits on hold, and brings back a transcript and summary.

Great addition. Time from missed call to booked slot is probably the real leading indicator, since a missed caller cools off fast. Tracking both together shows whether the agent is actually winning the job or just logging it.

Ha, agreed. The gutter negotiations can stay human for a while. Thanks for the thoughtful back-and-forth, Florian.

Jev works great for realtime speech classification, with the result ready by end of turn

Agree on the price line. A good setup lets the bot answer, qualify and book a slot, and hands anything that sounds like a quote or a commitment to a human. Bots negotiating gutters with each other is coming faster than anyone's ready for though.

Built one for phone calls (Patch). What sticks for non-devs is the boring errands: checking opening hours or stock, asking about appointment availability, following up with a supplier. The agent calls from your own number and hands back a transcript and summary.

Agree, disclosure should be the default. With Patch, the agent opens by saying who it's calling for and why, says the call is recorded, and confirms it's an AI if asked. If the person doesn't want to continue, it hangs up.

Fair point, and that's the right metric. A text that only says "we'll call you back" is just a slower missed call. Booked jobs per 100 missed calls cuts through the vanity numbers.

That's a nice design. Splitting schema changes from description-only changes makes sense, since a quietly reworded description can steer an agent just as much as a new parameter. Hashes keep the lockfile clean too.

Locking approved MCP servers and failing CI when one changes is a really smart default. Tool descriptions drifting quietly is the scary part. Does it diff the tool schemas too, or just the package version?

Missed calls are such a good wedge for small businesses, every one is a lost job. Are you texting the caller back first, or having the agent call them back?

Agreed, the summary is the real product. The useful version isn't "call completed," it's a clear outcome: what was actually agreed, what's still pending, and what needs a follow-up.

Completely agree. Suppliers, banks, clinics: so much back office work still lives behind a phone line. That's why we built Patch as an MCP server. Your agent calls from your own verified number, can work through phone menus and wait on hold, then brings back a transcript and summary.

Thanks for sharing the spread, the p90 gap is the useful part. A 2 second tail is exactly where callers start talking over the agent. Did you also measure how fast each one recovers after a barge-in, or just time to first audio?