Yeah, so I basically just wanted to like, we're starting to just still work in progress, but we're starting to like try to start to think about like what the top priorities from people that we've spoken to.

I think we have definitely taken into account what the conversation that you and I had last time that was very clear, like what your top like pain is around transcription.

Switch the text.

And so I'm just going to show this one sec.

Okay.

Can you see my screen?

Yeah.

Okay.

Cool.

So.

I basically just wanted to get your reaction to these and just see if it fares with you.

And then we could start to also just put down the biggest pains that you have,

if that's okay.

I mean, I can even add to the top one there already, especially on transcription.

There's more nuance to it.

I think it's

what's not on there right now is

transcription of low quality audio.

So especially like telephony, the 8,000 Hertz

Mula crappy audio quality that comes in over the phone.

That has been our huge problem right there.

Because models are typically trained on 16,000 or 24,000 Hertz.

And it's, it's, it's, they just do not do a good job of, of understanding, that kind of low quality audio.

Uh, we've even considered like doing our own like fine tuning, on all over a lower level quality.

Um, but it's, it still has not been,

has not been great, although we have found kind of a way around this.

Uh, and actually some customers prefer it now.

Uh, we've been using, since we're like, Our app is currently like multimodal.

So like it's, you can talk to it over phone, WhatsApp, texts, and it's all, you're all talking to the same agent.

And what we started doing is people actually liked doing WhatsApp voice memos.

And that quality is great because it's coming straight from their phone and it's going over WhatsApp and that's like high quality.

So like the transcription is perfect now.

But, you know, anyone who's relying on phone calls, forget about it.

Interesting.

Yeah, customers are loving the WhatsApp voice memos.

They find it actually easier than having a phone call.

Really?

Because then they can just read the response when it comes in.

And.

A lot of our customers, they don't really have even time for a call sometimes.

They're kind of multitasking so much that they just need to quickly pop off, like, all right, my agent just sent something and now I gotta respond real quick, then I gotta handle something else in the real world real quick.

So that's actually been amazing for us because

the low quality voice problem has been a real thorn in our side, clearly, as it was

my first pain point.

Okay, that's super, super interesting.

But by the way, do people, because someone else told me about that they were going to start using WhatsApp as well.

Do you not, like, because I know right here everyone uses WhatsApp, I use WhatsApp.

Yeah.

But my friends in the States, like, typically nobody uses it.

Yeah.

So this pretty much only applies to our customers in, you know, EU or UK.

So, unfortunately, US.

Yeah.

Yeah.

And the only reason I have WhatsApp is because I just know so many people over there that like basically bullied me into getting WhatsApp.

Yeah.

Deliverer.

Exactly.

Yeah.

Okay.

Very helpful.

Okay.

And then so how do you feel about the kind of like it's hard to reliably know how well conversations are going, what changes, what differences should chapters make?

Yeah, we actually started, I had one of my engineers actually working on it this month because we've been noticing this a lot with some customers where they're like, hey, our customers want to know when a conversation went poorly.

They want to know

they want to be floated up to them so they can intervene.

Um, so we haven't, ruled it out just yet.

Like we're real, like literally he merged it in today.

So we have not seen, what the, I mean, I'm a little, skeptical that LLMs can positively identify when a conversation has gone badly.

Or actually, you know what?

I'm more worried about false, false positives on that.

We'll see.

Uh, I, I trust my employee on this one.

But I have just doubts on the ability to determine that.

I feel like a conversation will mostly go 100% correct, but then there was maybe one mess up and it's going to consider the whole conversation messed up.

I've worked with LLMs enough to know that that's 10s what would it do.

I'm not sure that's what you meant.

Yeah, yeah, yeah.

Like if it was listening to our call, it'd be like, oh, this, you know, Jack said something and then, yeah, yeah.

And it just like misinterprets it or reads too much into something.

Interesting.

Like to kind of, what's the word?

Like human, some humans are like us as well, right?

Like overthinkers or something.

Yeah, yeah.

So we're trying to focus mostly on like, hey, did the conversation resolve?

Like in the end?

Did they get what they needed to go?

And even though if there was some tribulation along the way, but maybe it could have like varying degrees of like, all right, this conversation went sort of okay.

Yeah.

Like it's definitely been enough of a problem because call quality has been not that great that customers are starting to complain about this and they want to know when something's gone wrong.

Yeah.

So I mean, that alone is an indicator that Voice agents are not quite there yet.

Yeah, yeah, yeah.

Although there's sometimes like, it also depends.

Sometimes it's a problem of transcription, but sometimes it's also like the AI just has no idea based on it was like one instance was

oh yeah, like a customer called in and asked something that the AI had no way of knowing.

It was almost like a current thing like oh, hey, it's raining today.

Is the site open?

And in the AI was just like,

well, I don't really know.

And then it fortunately made something up and that's like, oh no.

Yeah, that's not cool.

So there's these hallucinations that can happen in there as well.

And again, that's something that picking up if this conversation went right or wrong.

Wrong, it's kind of hard to tell because it sounded like it was confidently, I mean, it was confidently incorrect.

So

do you have like, you know, I was thinking the other day this is just like massive aside, but like, like how we have like century when like, we hit like, you know, I don't know, there's just some sort of permissions error or something like API key expires.

We kind of learn immediately like shit, like we need to fix this.

Something's messed up.

Yeah.

That's actually a really good idea to create like custom alerts for that.

Like, we have like something similar for that, for a couple of things.

Like, if somebody, if like a, a number gets a lot of like dead calls, like people calling in and just saying nothing, we have alerts for that right now.

But this is actually a good idea to like have alerts for like poor conversations happening.

Yeah, I guess it's like what you said, like that balance of like, not, if you have too many false positives, it's just like, You're never going to even check the stream notifications, but you don't want to miss the actual ones where it's told someone, oh yeah, it's open, drive over three hours.

Terrible.

Exactly.

Okay, super interesting.

And this was kind of also,

how do you feel about number three, turn-taking, break, conversational flow?

I would say it's still a problem for sure, but I think definitely not as much as the other ones.

I think what we started doing is just allowing more time for the agent to pause and think about something so it has more time to get the correct transcription and really know when the person stops speaking.

And the one thing that we use for that is just having a little waiting noise.

Like, it's an eye thinking noise.

We're just like a,

you know,

something like that.

And that, that go, that goes hard, you know, that, that people, people know to wait for that.

So good.

So good.

It's that classic that, you know, that whole, like, elevator thing where they, like, people complain, you know, that story is, like, classic design thing of, like, slow elevators.

Put in mirrors and people stop complaining.

Yeah.

Yeah, that's great.

One side thing, just that like, I haven't got this on there, but like some people I hadn't mentioned, do you have any challenges around onboarding?

Tons, I would say.

We ended up like hand holding a lot of our customers because of this.

Or like they don't even want to try it unless, like they don't even want to demo it unless we completely set it up for them.

And then we give them a number and then they like, we'll try it out.

And then that works really well.

But getting them to, we have not had many

people coming in from the cold and setting them up themselves and getting running up and running, like basically zero at this point.

Okay.

So you're spending a lot of time on boarding.

Yeah.

A lot, a lot of time on boarding, unfortunately.

And, and.

I think it's been better because I think we would rather handhold them and make sure that the agent is set up correctly than have them try to do it themselves and then have a poor experience and then just bounce.

Yeah, but obviously that's not ideal.

Obviously we would love it to create our own.

We've made so many adjustments to our onboarding process at this point where we even try to make it more automated, have AI, you just give them a basic, like you give them your, if you have like a, let's say you're a restaurant, you just give them your, your website and then we'll do all the research in the back and like the AI will do the research in the background and try to set up the agent for you as like some first pass and then you can give feedback.

But even they don't want to do that.

You know, it's, I don't know if people are just like already getting too used to AI companies and they're like, hey, this all should just be done for me.

Yeah.

Yeah.

But it's been tough.

Been tough.

Interesting.

Yeah.

This is how do you think that if you're slotting that into like pains, like do you think onboarding makes it into top three or is it like not in top three?

You know what?

I would say that that makes it to top three.

I would replace the taking turns at this point and say like the onboarding is up there.

But below, not as big of a problem as the other two.

Yeah, I would say not as much.

Just mostly because we're still, since we're trying to mostly target more high valuable customers anyway, it's not that big a deal.

We're not too worried about that as much.

But I mean, it's something we would love to have automated if we want to scale.

Yeah, I have.

I mean, the others that you can see how it might be like, technology could maybe unlock it, but it's like, that one feels like it'll be tough, but I don't know.

Do you think there's like technology that could sort of make onboarding customers easier?

Or do you think it's just like a human challenge?

Yeah.

I mean, theoretically, you would think AI can unlock this somehow, but it would take so much effort to do so.

It would almost be like a whole new company we need to spin up just to do this.

Yeah.

Well, we're trying to solve problems in this space.

Yeah.

I don't know.

It's come up a couple of times now.

I mean, it's a big one.

Yeah.

I mean, if you're seeing a lot of other people having this issue, it's definitely something to it.

And it also feel good to know that we're not the only ones that are having this onboarding issue.

Yeah.

Some people say integrations, spending a lot of time on integrations.

Yeah.

Stuff like that.

Yeah.

So, well, that's definitely one thing everybody wants.

It's like if you're trying to target real estate, then it's like, well, I have this real estate software.

Can you integrate with that?

Oh, great.

It's like, does it have an API?

And then you receive a PDF.

Yeah.

Well, maybe everyone will just have NCP servers running at some point and we can just all talk to each other and integration should be easier, but it's probably gonna be a ways for that.

I hope so.

Yeah, I don't know.

We'll see, I guess.

Okay.

Amazing.

That's super helpful.

So we're probably only imagine I have time to get to gains, but.

Kind.

Of is like the sort of like the, you know, the.

Result of that problem.

Like, if you had, you know, if you, if you were to have solved that, what you would get out of it, basically.

So,

like, perfect transcript.

Like, you know, if you're imagining, like, something, the wave in the magic wand, like, that you would get out of it, like, so we've got here, like, perfect, trans, perfect, quote, unquote, transcription, maybe say, like, on low quality audio or something.

Conversations feel instant and natural.

I feel like maybe I can actually just like just put the conversations feel natural.

Yeah.

And know how every conversation is doing, I guess.

Being alerted.

Being alerted.

Yeah.

And I guess if we add another one, like customers can onboard themselves.

Yeah.

Self serve.

It's the dream.

Yeah.

You know, I think a big problem with the onboarding process is that because this is such a new kind of technology, people are just almost scared.

To try to onboard themselves because they're worried about, they don't know what the capabilities of are.

I think they're more worried about messing it up and not being correct, that they just kind of throw their hands up and rather just have somebody help them.

Which again, AI should theoretically help.

Yeah.

I think I'm also, you probably see this a lot in this space is, you know, AI companies tend to think AI is the solution to everything.

So it's like our only tool in the toolbox.

They're like, oh, I'll just throw AI at it.

You know, that'll fix it.

Yeah.

It's not that simple.

Yeah.

It is interesting.

I don't know.

It's always like, yeah.

I don't know how, I don't know, like, yeah,

I just had a deal with like a broadband company today and like, you know, they still have humans and it doesn't feel like it feels like it's like a lot of it's like the negotiation and things like that.

And it's like, can you, I mean, you can get AI that does the negotiations.

Can you get the AI to understand I don't know.

Like, is it.

It gets into, like, interesting.

I don't know.

And I guess that's a form of.

I don't know if that councils onboarding where they send me, like, I still prefer to onboard over the phone than

just like, I could have done it online, but I knew I'm gonna get a better deal on the phone.

That's very true.

Oh.

I miss if that ever goes away.

If you're just getting the same raw deal as going online.

You're not talking to a human anymore, you know?

Yeah, I don't know.

Yeah.

Okay.

And so, sorry, Anthony.

So do you think that these would, like, is this how you would, if you were thinking of what you would like to just exist?

Is this natural order that you think?

Yeah, I think so.

Okay.

Okay, cool.

Super helpful.

Is there anything that isn't on there that should be on there?

Nothing I could think of right now.

Okay, cool.

Thanks.

That's really helpful.

Yeah, so this is like, I've honestly found these last two ones like, I haven't I've been doing it as much, to be honest, verifying because it feels like these are the two main ones.

But.

Yeah.

Do any of these resonate or not resonate?

It's just too fancy.

Yeah.

Orchestrated and then calls.

Yeah, I think.

I don't know.

Like what the.

Yeah.

I don't know, transport network like of the, yeah.

No, I mean, like, just like the logistics of setting all that stuff.

Yeah, just like the managing all like the web sockets and everything and just all that sort of like sending the audio, receiving audio, like all that sort of stuff.

Which is the part I found the most fun while building stuff.

But I like that kind of stuff.

Yeah.

It is fun.

It's fun.

That's what I'm doing right now, actually.

So, yeah.

I was just reading the second one, the instrument system with recordings, eval sets, latency scoring.

I mean, ideal.

Yeah.

I feel like we end up just kind of like, getting around that by being like, okay, like in the end product, is it, you know, actually, that's, that's, I guess that's not a good way to go about that, too.

Like, you know, determining after the call, if the call went poorly and then doing something about it, then, you know, it's more firefighting than being proactive.

Catching a jury, getting someone.

Yeah.

Although we, we do have.

A lot.

Of, and this may be more like, is each individual system like, hey, did our text of speech start being slower or our transcription started being slower and switching over to a different provider at the last second so that we know this.

We actually do have that right now.

Yeah, we had to implement that because

certain services like Deepgram can be quite unreliable at times.

Their transcription quality is great, but their latency like, it'll.

They definitely hit, like, some limits at times, and, you know, it's sometimes faster to just switch back over to, like, 11 Labs' transcription or their Texas Beach, depending on, like, 11 Labs tends to be more consistent on their latency from what we've seen.

That's super cool.

That's really interesting.

It's interesting that you managed to.

That's great.

Yeah.

You measure just the response on the request basically and then change it once.

It hits a certain threshold over a certain amount of time.

That's super cool.

I realize we're coming towards the end.

Oh, wow.

Oh, wow.

Yeah, that went fast.

Thank you.

That was extremely helpful.

Yeah.

So we're going to have some other stuff to share hopefully next time if it's still okay.

Absolutely.

If there's anything that we can help with as well.

Please let us know.

I don't know if, like, we're gonna set up swag and stuff soon.

There's.

There's just been tons of debate in the team around, like, water bottle choices and all this sort of stuff.

So I've kind of checked out of that, but, like, someone's debating it.

Oh, yeah, yeah.

You can spend so much of your.

Your time thinking about swag, you know?

Before you lose your mind.

Yeah, exactly.

But, yeah, Anthony, thank you so much.

I really appreciate it.

I know your time is very precious right now, so thank you so much.

And we're trying to make it worth your while.

Thank you very much.

No problem.

It's a lot of fun.

Thanks.

See you later.

All right.

Bye.