Meet Grace: Sidecar’s New AI Agent Takes the Stage | [Sidecar Sync Episode 126]

Summary:

 In this episode of Sidecar Sync, Amith Nagarajan and Mallory Mejias pull back the curtain on one of Sidecar’s most ambitious experiments yet: Grace, a real-time AI audio agent designed to transform how organizations engage with users. From a beachside hackathon to a live deployment in under two months, they unpack the full journey—covering the business case, rapid development process, tech stack decisions, and early learnings. Along the way, they explore how audio AI captures nuance in ways text cannot, why “progress over perfection” is a critical mindset for innovation, and what this means for associations looking to modernize their member experience. If you’ve been curious about voice agents, AI-driven sales funnels, or how to actually build with AI (not just talk about it), this episode is a must-listen.

Timestamps:

  🙋‍♀️ Chat with Grace on the Sidecar website

https://sidecar.ai/ 

👥Provide comprehensive AI education for your team

https://learn.sidecar.ai/teams

📅 Register for digitalNow 2026:

https://digitalnow.sidecar.ai/digitalnow

🤖 Join the AI Mastermind:

https://sidecar.ai/association-ai-mas...

🎀 Use code AIPOD50 for $50 off your Association AI Professional (AAiP) certification

https://learn.sidecar.ai/

📕 Download ‘Ascend 3rd Edition: Unlocking the Power of AI for Associations’ for FREE

https://sidecar.ai/ai

🛠 AI Tools and Resources Mentioned in This Episode:

 ElevenLabs ➔ https://elevenlabs.io/

Claude Haiku 4.5 (Anthropic) ➔ https://www.anthropic.com

Google Nano Banana 2 ➔ https://deepmind.google 

👍Please Like & Subscribe!

https://www.linkedin.com/company/sidecar-global

https://twitter.com/sidecarglobal

https://www.youtube.com/@SidecarSync

Follow Sidecar on LinkedIn

⚙️ Other Resources from Sidecar: 

• Sidecar Blog
• Sidecar Community
• digitalNow Conference
• Upcoming Webinars and Events
• Association AI Mastermind Group

More about Your Hosts:

Amith Nagarajan is the Chairman of Blue Cypress 🔗 https://BlueCypress.io, a family of purpose-driven companies and proud practitioners of Conscious Capitalism. The Blue Cypress companies focus on helping associations, non-profits, and other purpose-driven organizations achieve long-term success. Amith is also an active early-stage investor in B2B SaaS companies. He’s had the good fortune of nearly three decades of success as an entrepreneur and enjoys helping others in their journey.

📣 Follow Amith on LinkedIn:
https://linkedin.com/amithnagarajan

Mallory Mejias is passionate about creating opportunities for association professionals to learn, grow, and better serve their members using artificial intelligence. She enjoys blending creativity and innovation to produce fresh, meaningful content for the association space.

📣 Follow Mallory on Linkedin:
https://linkedin.com/mallorymejias

Read the Transcript

🤖 Please note this transcript was generated using (you guessed it) AI, so please excuse any errors 🤖

[00:00:00:14 - 00:00:09:17]
Amith
 Welcome to the Sidecar Sync Podcast, your home for all things innovation, artificial intelligence and associations.

[00:00:26:03 - 00:00:27:19]
Amith
 My name is Amith Nagarajan.

[00:00:27:19 - 00:00:29:16]
Mallory
 My name is Mallory Mejias.

[00:00:29:16 - 00:00:41:02]
Amith
 And we are your hosts. Actually we are two of your hosts and there will be a third co-host joining us very soon. But I don't want to give away that surprise Mallory just yet. How are you doing today?

[00:00:41:02 - 00:01:06:14]
Mallory
 Oh, Amith. Yes, I'm excited for our third co-host to join. Again, no spoilers but stay tuned for that. I'm doing well still in the thick of this house project. That seems never ending. We are officially moved in. Our house is chaos. Still no kitchen. Furniture is everywhere. Box is everywhere. But I'm glad to be able to sit down at my desk and have a nice Sidecar Sync episode recording. I feel like that really grounds me in the week.

[00:01:06:14 - 00:01:08:18]
Amith
 That's awesome and you have internet too.

[00:01:08:18 - 00:01:40:17]
Mallory
 Oh, speaking of internet. I told you this, Amith. But everything was set up. Internet was one of the first things before we moved in. I was like, I've got to get internet set. I work from home. You know, I use the internet for acting stuff as well. And about a day before we moved in, our contractor severed our fiber wire for internet. And so I was panicked. I was calling AT&T. But thankfully, they actually, if this happens to you, it's apparently a very easy fix. I thought it was going to be the world's most difficult thing to do. But they came in, put a new fiber wire and we were good to go.

[00:01:40:17 - 00:02:33:19]
Amith
 That's awesome. Well, fiber is amazing. Having fiber optics to the home is just, we're definitely living in the future. Well, that's awesome. I just got back from a really cool trip to Washington, D.C. Went there to see some family, but then also hung out for an extra day and met with a whole bunch of different association leaders. So I always find that energizing because I always learned so much just talking to colleagues who were in associations, doing the work of advancing their missions and their organizations and the things they're challenged with, the things they're excited by. So I try to spend a lot of my time when I'm traveling, spending face to face time along with associations, CEOs and other leaders. So that was an awesome time. And now I'm back in New Orleans where it's 40 degrees today and New Orleanians are running around in, they're not really running around, they're scarcely found on the street. But to the extent they are, they're wearing their parkas and their winter hats and so forth. So it's quite fun.

[00:02:33:19 - 00:02:51:18]
Mallory
 It's about a brisk 38 here in Atlanta. So a little chilly, but I'm hoping this is the last cold front maybe before we get into some nice spring weather. But Amith, I wanted to ask you, meeting with these association CEOs, are there any interesting AI projects that you heard about that you could maybe give us a high level overview of?

[00:02:51:18 - 00:03:07:15]
Amith
 Yeah, I mean, you know, the spectrum is wide. And so I met with an individual who is leading kind of a good sized association about 100 plus staff. And he is focused on a really deep transformation this year, rescaling his entire team top to bottom,

[00:03:08:24 - 00:03:15:22]
Amith
 also pursuing the AIP for basically every staff member. They're also in the process of three or four pretty major projects.

[00:03:16:23 - 00:06:03:16]
Amith
 And you know, transformative stuff. I mean, things that are not just about making their internal operations more efficient, but really remaking the experience of the member. And that's so key and really relevant to what we're talking about today as well. But the idea that AI can remake how we do our work inside the house is important. But how we welcome people into the house, how we have the members interface with us is so incredibly critical, especially in a world with a member's experience as just as a human in the world today. Our experiences are shaped so actively by leading consumer brands that are redefining really what the expectations should be. And that applies to all of us as associations. Association leaders need to not just get better than they were last year, but they need to keep pace with the general experience of consumers. And whether we like it or not, that's the expectation. And the good news is that there's no disadvantage to being an association anymore. It used to be that associations, because they're way smaller than large commercial enterprises and don't have the technology resources, et cetera, et cetera, always kind of had a mixture of this, you know, kind of a, I would say not quite a defeatist mindset, but the view that they just couldn't possibly keep up with the Amazons and the Netflixes of the world. And there was some rational thinking, of course, that went behind that when, you know, the critical mass of financial resources was required to do what we're talking about. But that's democratized now. You know, AI is available for very little money, honestly, now, and you can implement it in ways that you probably don't realize. And so Blue Cypress, our BHAG at the top level companies, the sidecar is one of the businesses within Blue Cypress. And Blue Cypress has a BHAG that is to help associations be as powerful as the Fortune 500. And when we talk about that, we're specifically referencing this inflection point where power is no longer by aggregating financial resources, but power is actually about just being smarter about how you use what's available to you and the abundance of AIs, of course, a key piece. So coming back to your question, that was one meeting. I had another meeting with a person who was very early in the journey, who was, you know, very excited, really, really interested in doing things, but hadn't yet really embarked too far. So we were just talking about the basics of, hey, how do you put in place a solid, you know, initial roadmap? How do you make sure that you're being safe in the way you're using data with AI, stuff like that? So I think associations are, you know, of course, at various stages of their journey. And I found each of those conversations equally compelling, you know, and others I had during the day where it's not that you're at the forefront of it, but that you're really trying hard and you're pushing yourself and your organization to adapt and learn. That's what I find interesting. And I think people who are out there making that investment in themselves and their team, they're going to see incredible results.

[00:06:03:16 - 00:06:42:10]
Mallory
 Well, Amitha, I think it's time to segue into our topic for today, which is going to be a little bit different from our usual format. For those of you who have visited Sidecar.ai recently, our website, you might have noticed a new addition, an AI audio agent named Grace. Grace is a conversational voice agent that lives on the Sidecar website. You can talk to her about Sidecar offerings, ask questions about our AI learning hub, and she can walk you through things like purchasing for yourself or for your team. She also pulls up visuals as she's talking you through things. She's brand new. She's been live on the site, Amitha, for how long? I was going to say less than a month, but even less than that.

[00:06:42:10 - 00:06:51:17]
Amith
 No, much less than that. I mean, we've had Grace in test mode internally and with just a handful of good friends of the company for maybe two weeks.

[00:06:53:07 - 00:07:03:13]
Amith
 Grace has been a development for two months. For the last two weeks, we've been actively testing with external folks. And then really, Grace went live on Monday of this week. So I think it was yesterday.

[00:07:03:13 - 00:07:36:18]
Mallory
 Okay. So way less than a month, basically less than a week, just a few days. And what we want to do in this segment is walk through the project, Amitha and myself and Grace, from the ground up, the way any association leader would want to understand it if they were thinking about building something like this for their own organization. We're going to cover the business case, the build, the trade offs, the tech stack, and what we're learning from very early deployment. But we figured instead of just talking about Grace on the pod, it would be better to have her join us ourselves. So Amitha, can you kick that off for us?

[00:17:16:03 - 00:17:31:19]
Mallory
 Amitha, what was most exciting for you in that conversation with Grace? I'm sure you've had some other conversations with Grace, but was there anything that surprised you as one of the leaders of this initiative, the builders of the initiative that Grace could do?

[00:17:31:19 - 00:17:36:14]
Amith
 Well, Grace and I are good buddies. We've been having lots of conversations for the last several weeks.

[00:17:38:02 - 00:18:08:08]
Amith
 I'm just amazed by how good Grace is at picking up on the nuance of conversation, even sensing hesitation, sensing different emotions, essentially. That's perhaps one of the greatest gifts of audio AI is the modality conveys a lot more information. We've talked on this pod about how when you take audio and you convert it just to a text transcript, you lose a lot. You lose a lot of information. You lose the tone, the volume, the inflections.

[00:18:09:13 - 00:19:01:07]
Amith
 Audio AI, on the other hand, does not. Audio AI picks up on all of that. Grace is a great example of that. Still a very early example of what will soon become far more sophisticated audio AI, but Grace is quite adept at understanding the person she's talking to. I've had lots of simulated calls with her, actual calls with her. I'm trying to simulate being different kinds of personas, very aggressive, very forward-looking, kind of person versus someone who's very meek and very concerned and very worried and people who have deep expertise in AI who are kind of testing what she knows and people who know nothing about AI and everybody in between, people who are executives, people who are entry-level. I've tried a bunch of different things. Grace is not only able to handle a lot of these conversations really well, but I think picks up on these nuances. That to me is what's most exciting about this modality in general and Grace in particular.

[00:19:01:07 - 00:19:47:21]
Mallory
 I also want to point out something too. Amith, you were asking her about the Teams AI Learning Hub and the phrasing you used was, "I think I want to get my whole team in the mix." Going back to nuance of conversation, the fact that she immediately was like, "Oh, get them in the mix. That means you're interested in the Teams AI Learning Hub for your whole organization." There were so many moments like that where I am just mind-blown. We talk about this stuff every week on the pod. Totally. The visuals are incredible. The checkout experience at the end is insane, which we're going to talk about all of that. I want to dive all the way back to two months ago, not that long ago. What problem was Grace built to solve? What was the gap in the sidecar experience on our website that made you say, "Hmm, we might need a voice agent here."

[00:19:48:21 - 00:20:07:22]
Amith
 Well, those of you that know me or have heard me speak about this on the pod in the past, I'm a big fan of this concept called hackathons. A hackathon is nothing but a dedicated focus time period where you take a group of people and preferably in person, you isolate yourselves from the rest of the world and from other activities and priorities and you just focus on something.

[00:20:09:09 - 00:21:11:07]
Amith
 At Blue Cypress, we tend to run these hackathons at least a couple times a year for software development. We run hackathons actually on other things. We've had them for marketing and sales and for finance and for all sorts of different functions. What we did in January, at the end of January, we got a group of 14 people together down at the beach in Florida and we essentially worked for 16 to 18 hours a day, six days in a row. That sounds really miserable to some people, but it was extremely fun. It was mostly self-directed and what we said coming in was, "Hey, this is going to be the audio AI hackathon." I started the event by just explaining to people why I felt audio AI was going to achieve its chat sheet DT moment, if you will, sometime this year. It's this confluence of the power of AI models being sufficiently strong to be useful. The speed of the AI models being good enough for really high quality real-time conversation, as you just experienced with the GRACE demo we just did, and the nuance and the ability for the AI models to really adapt.

[00:21:12:16 - 00:21:55:16]
Amith
 We started off not with the idea of building GRACE and actually many things came out of the hackathon, not just this project. We came up with, "Look, all 14 of us just go experiment with audio AI in general." Some people did stuff with 11 labs, which is the technology that GRACE has partly built on. Some people worked with local AI models for text-to-speech or speech-to-text. People did a whole bunch of really interesting creative stuff. Without, and this is a key part of the hackathon, without the pressure of creating something for any durable purpose. It's not that we said we're going to create a product, but we suspected probably some ideas would be product worthy or things that could turn into products, but that wasn't the goal. The goal was just to experiment and really be in a sandbox and play.

[00:21:56:21 - 00:21:57:16]
Amith
 That's super fun.

[00:21:59:03 - 00:22:13:16]
Amith
 One of the developers, actually one of our earliest career folks, had come into Florida. He actually had to fly in overnight from the West Coast and his flight was canceled. This was late January. There was a whole bunch of weather in the area and the South.

[00:22:14:18 - 00:22:38:05]
Amith
 He finally got there. Right on the plane, he built a portion of what became GRACE. It's a good thing. Thankful for caffeine and young people's energy because he came into the hackathon with an amazing demo. Part of what we do in GRACE is we utilize a piece of 11 labs as technology, which allows you to do real-time interactivity in the browser.

[00:22:39:09 - 00:22:57:16]
Amith
 The AI is smart enough to say, "Hey, I'd like to do something called a client-side tool," which essentially is on the browser being able to do something in the browser. What we decided to do with it was to display different kinds of images that share richer information than just the audio alone can convey. You just saw that in the demo.

[00:22:58:18 - 00:23:20:16]
Amith
 What this individual did is actually he built a demo computer store, where it wasn't like a fully functioning live store, but he showed how he could talk to an audio AI to browse, search, and ultimately buy a laptop with his demo. That opened up a lot of ideas from different people. I personally had no idea that 11 labs had this concept of client tools,

[00:23:21:23 - 00:23:57:02]
Amith
 and that opened up a whole bunch of creativity for the rest of the group for that whole week. The goal initially in January was, "Hey, we're going to do a bunch of AI stuff with audio specifically," and then out of that, then we said, "Hey, this is going to be a really big deal. Let's make an audio AI agent to solve a key problem for both Sidecar and its prospective customers." That was designed not to take the place of, but to supplement the human sales team that Sidecar has who only work 12 hours a day or whatever they're doing. They're working real hard, but they're not there 24, 7, 365.

[00:23:58:24 - 00:25:10:07]
Amith
 Even if they were, they couldn't keep up with the demand. We're very fortunate that Sidecar's growth rate has accelerated along with AI, and we have a lot of people very interested in learning AI in the association space. People are coming to our website all the time, trying to learn about the certification, trying to learn about our courses. We said, "What if we had a really awesome AI audio guide that could help people understand if it's the right fit for them, walk them through the different options, talk not just about AIP, but if they're interested in the mastermind or interested in our events like Digital Now or the Innovation Hub, someone who really knew the Sidecar product suite really, really well, like our sales team would. That's essentially what Grace was built to do. Again, not to replace that, but to supplement it. Our hope is that this achieves two goals. One is we want to provide our clients and prospective clients better service. The other thing is, and better service not only means 24-7, but it means in the modality they prefer. Some people, myself, honestly included, I oftentimes don't like booking costs with sales people because I just don't want the perceived pressure, whether it's perceived or real. I just don't want to be on the hook to talk to a salesperson. Talking to AIs feels a lot different. I can just hang up on them when I walk to them.

[00:25:11:19 - 00:25:19:01]
Amith
 That's kind of the thought process initially that went into it. The idea for Grace came out of that, and here we are six weeks, seven weeks later, and we have a lot of products in the market.

[00:25:20:05 - 00:25:47:09]
Mallory
 I'm with you as well. I do not like booking sales calls, especially. It's like a dance almost, and they don't give you the pricing. They want to wait until the second call to give you the pricing based on whatever you need. I too would opt to talk to Grace on a website if she was there and I was interested in purchasing something. It sounds like we're supporting our sidecar team because they can't work 24-7, 365, and also supporting our customer or prospective customer experience as well through Grace. Is that right, Ami?

[00:25:47:09 - 00:27:21:04]
Amith
 Yeah, definitely. We think of Grace as being, if you think about the sales funnel in a classical sense, where Top of Funnel is like earliest visitors to a website. They're just getting awareness to the brand, understanding what the thing is. Mid-funnel where someone's really starting to kind of qualify what it is that you do and how good of a fit you are. Bottom of Funnel where people are really interested in purchasing and they're just trying to get through that purchasing process, we wanted Grace to help at all three phases of the funnel and to be really a full funnel inbound sales agent essentially. Grace doesn't call you. You choose when to contact Grace, but with Grace, you can learn from scratch or as you saw in the demo, you can actually check out and do a fairly sophisticated B2B level transaction where you're purchasing a subscription for all team. You can do that for an individual. We plugged in Stripe to make it super simple. If you've ever used any website powered by Stripe, your information is pre-populated. It's all totally secure that way too. It's just really easy. So anyway, I think with Grace, our goal was to reduce friction, make great information available, and hopefully put people at ease too. That was one set of objectives is the sidecar experience. But the other thing that we're trying to do is we as sidecar, we want to live in this glass house that's always experimenting, always pushing the envelope, always trying new things because we see that as our role. Sidecar in many ways is very association-esque. We have subscription products. We deliver education. We have events. We have content. We do a lot of association-type activities.

[00:27:22:08 - 00:27:39:06]
Amith
 If Grace can be an effective type of AI tool for sidecar, perhaps it means that technology like Grace, Grace herself or other technologies like Grace, might be useful for associations. Maybe we'll open up some creative thinking for some of our friends in the association community. So that's also the hope as well.

[00:27:40:08 - 00:28:08:24]
Mallory
 I want to talk a little bit more now about the build and the technology behind it. So as you mentioned, Ameth, you moved incredibly quickly and the team moved incredibly quickly. Having this built in 60 days or less, you made a post in our Blue Cybers Teams channel earlier that this really exemplified our core value of progress over perfection, which I love. So can you talk about the process of moving so quickly, what that looked like and how many people were involved? Would something like this be realistic for an association? What do you think?

[00:28:08:24 - 00:28:13:07]
Amith
 I think the speed that we moved at would be pretty difficult for a lot of organizations.

[00:28:14:13 - 00:28:36:13]
Amith
 But we're not unique in that sense. What we did is we allocated some dedicated resources. We cleared the pathway for them to experiment and we iterated really rapidly. We're also tolerant of some level of risk here, because Grace is not perfect. In fact, in the demo, you may have noticed that she didn't pronounce my name perfectly. She did a decent job. She got your name right.

[00:28:37:14 - 00:28:43:03]
Amith
 She did not get AIP correct. She said AAP once, I think, and then something a little bit off.

[00:28:43:03 - 00:28:44:02]
Mallory
 You struggled with that one.

[00:28:44:02 - 00:29:12:16]
Amith
 There were struggles with that. Yeah. I mean, the audio is not perfect, but that's actually precisely why I think it's important to experiment out in the open with this stuff right now, because people actually make the assumption that audio AI is not capable at all right now. In fact, I've shared the Grace link with people before we went live on the public website, which by the way, if you just go to the Sidecar website, which is sidecar.ai, on the top right corner, you'll see the Grace icon. You can click on it. Grace adapts to your mobile browser as well if you're on your phone.

[00:29:13:22 - 00:31:26:15]
Amith
 Grace will show you images in a mobile-oriented format too. But the idea essentially is when we think about this stuff, we look at it from the viewpoint of risk tolerance. Progress over perfection, what it's essentially saying is perfection is this ideal that actually is not perfect, in that it actually requires you to effectively guarantee that you're not going to ever get there if you're trying to shoot for perfection. All too often, I see organizations, associations included, in fact, very much included in this, that only settle for this must be perfect. This must be our incredibly high quality standard. And there's cases where that makes absolute sense. If you are setting standards for surgical procedures, you better make sure those things are perfect, as perfect as humans can make things at all, right? There are things that are mission critical at that level. If you're putting people on an aircraft, you better make sure that you're as perfect as you possibly can be. But in certain cases like this, what's the worst that can happen? If someone has a bad experience of grace, hopefully they'll tell us about it. But we'll learn from it because as we say on the very top, it's a research preview and we record all the calls. We're going to learn from it. We're going to learn in the open and we're going to learn by experimentation. So our mindset is that we need to learn through process, through experimentation. And that's what our core value of progress over perfection embodies. You know, Mal, we've talked about core values a little bit on this podcast, but they're really the operating system for culture. And when you think about core values, it's when you differentiate core values from one company to the next, they really should tell you not what's good or bad, but really what's different about the company. And so for us, it's one of the core values that really embodies the mindset we want our team to continuously take is to focus on making progress, not to try to achieve perfection. That again, doesn't mean that we're not trying to achieve very close to perfection in certain things. And when we have, for example, production software that we're running in one of our companies like Rasta.io, for example, which processes tens of millions of emails a day for our customers, we're incredibly focused on near perfection in the way we operate that business and all of the others that are like that.

[00:31:27:18 - 00:31:46:16]
Amith
 Obviously, that's the case for like our finance department and things like that. But it's not the case for when we're out there experimenting with new products. That's really the mindset that I think is critical more so than technical skill, more so than having a lot of resources available. We're fortunate to have a good bit of all that, but it's more the mindset that allows us to be able to experiment this quickly.

[00:31:48:02 - 00:31:52:08]
Mallory
 How many people would you say were working on this over the last two months?

[00:31:52:08 - 00:32:09:04]
Amith
 On a dedicated basis, there was a couple of FTES on it, that's it. And then we had probably six other people that were involved fractionally, very fractionally, myself included. I was only involved at a very, very high level. Kind of my role tends to be two things. One is I'll help establish the initial vision and then I'll bark at people.

[00:32:09:04 - 00:32:12:05]
Mallory
 But I figured it was going to be one of those two things.

[00:32:12:05 - 00:32:28:15]
Amith
 Yeah, so that's basically what I do. So it seems to work pretty well. But I have a whole team of people that actually run the business well and they're driving the priorities and stuff like that. So I'd say probably six to eight people worked on it and a couple of people were heavily focused on it for the last two months.

[00:32:28:15 - 00:32:45:22]
Mallory
 Okay, so we have an inciting event, which was the hackathon. A creative moment doesn't have to be a full hackathon, but some dedicated deep work time, six to eight FTES with some other people fractionally involved. What AI model is powering Grace's brain and helping her generate these responses?

[00:32:47:03 - 00:32:48:11]
Amith
 We have another visitor for the pod.

[00:32:48:11 - 00:32:50:12]
Mallory
 Oh, and who is this?

[00:32:51:14 - 00:32:52:19]
Amith
 This is Lucy.

[00:32:55:07 - 00:32:57:21]
Mallory
 Where are my dogs? Lucy is adorable.

[00:32:59:00 - 00:33:00:19]
Amith
 Thank you. She's at the office with me today.

[00:33:02:07 - 00:33:13:07]
Amith
 And that's what Lucy is saying. She just decided she had to jump in my lap. Lucy is 11 months old. She's a Cavalier King Charles Spaniel and she has an older sister Penelope. She's a 12 year old Cavalier at home. So, um,

[00:33:13:07 - 00:33:18:12]
Mallory
 you guys don't be surprised if we have a Lucy AI product coming up near you soon.

[00:33:18:12 - 00:33:20:05]
Amith
 That's a great idea. That's a great idea.

[00:33:20:05 - 00:33:20:16]
Mallory
 That's adorable.

[00:33:22:07 - 00:35:57:00]
Amith
 But, uh, you know, when, uh, when we're thinking about, um, all of these different topics, kind of coming together and product development, um, the, the, the core focus for me is are we, are we doing something that's like driving any kind of differential change, like for, for us as an organization. And, um, I think that, um, when we look at it, the technology stack, you know, is, is really critically important. Is the, is the tech going to allow us to move quickly enough? Um, and in this particular case, um, we, we evaluated a whole bunch of different audio technologies. We ended up coming back to kind of an old favorite, um, which is a company called 11 labs. It's, uh, the number 11, but spelled out. So the word 11 labs.io is their website. They've got some great technology, um, a variety of different things they have. They have a creative suite that allows you to do a whole bunch of stuff, like, you know, create like, um, different kinds of synthetic voices that you can do text to speech. Uh, we talked about that in the past on this podcast and we, we, we think really highly of them. Well, about, I think about six months ago, they released a new product called their agent studio. So 11 labs agent studio is a voice agent technology that allows you to fairly easily spin up, um, audio agents like grace. Um, now there is a lot of work on top of just using 11 labs that we did in order to make a system, but 11 labs is the core, um, text to speech and speech to text pipelines. The way it works is grace is essentially, uh, taking, you know, voice prompts or text and turning them into audio. Um, and that's the text to speech and then grace listens. And so that's essentially capturing, you know, the waveforms of, of when we're speaking to grace, and then that converts from speech to text or STT. And then that speech, the text goes into a plain old LLM. In our case, we ended up using Claude Haiku 4.5. So if you're a user of anthropics, Claude, um, they have three different sizes of models in each of their model versions. They have Haiku, which is their smallest and fastest model. They have sonnet, which is kind of the mid-sized very intelligent, but kind of mid-sized model blend speed and performance. Um, and then Opus, which is their largest, most sophisticated model. It's slower, but far more intelligent. For the purposes of this project, we did not need Opus or sonnet. So we chose Haiku because of its speed. Um, and so we ultimately ended up with, um, you know, a, a model that was fast enough to be like a real time type of an experience for our customers. And so that was really important.

[00:35:57:00 - 00:36:05:19]
Mallory
 Okay. So grace is not processing the audio itself. She's turning the audio into text and then processing that.

[00:36:06:19 - 00:39:16:03]
Amith
 Yes. And the way 11 Labs works is it's able to actually capture a lot of the nuance that we were talking about earlier, where, um, you know, some of the insights about tone and, uh, intent and things like that are captured by the model as notes, then they're like kind of supplements the actual text and that helps the Haiku model understand what's going on now it's interesting because what we're currently doing with this, it's, it's basically called an audio to text pipeline, which is audio to text to audio. Um, it's very sophisticated as you've experienced now on the show. And I hope you all experiment with it live. Grace is quite good, but there is a gap in that because there's a translation happening from audio to text and back from text to audio. There are actually newer technologies out there that are straight up audio to audio models. In fact, there's something from Nvidia that was released back in January called persona plaques and videos persona plaques model is something that is not generally available from cloud providers, but looks very promising at some point. It or technologies derived from it or similar to it will be quite compelling. And the difference is you have full duplex audio to audio and that's kind of a lot. So let me break that down full duplex compared to half duplex means that the model can listen and speak at the same time. Now in the demo, you may have noticed I intentionally interrupted grace. And if you recall, it took a moment for grace to actually stop talking and listen to me. Um, and some of us humans are like that too. Um, I tend to be that way sometimes. Um, but, uh, you know, full duplex means that the model is capable of activating its ears and its mouth at the same time. So to speak, that's not what you experienced. Grace is half duplex because the model is not capable of both processing input and generating output at the same time. We kind of simulate the idea of full duplex by having the microphone always on. And when we detect someone speaking, we stop the audio model from emitting more audio tokens and we then regenerate. And so that's kind of the brief delay that you heard. Um, it works, but it's not optimal. So full duplex audio will be great. And then audio to audio simply means that the model itself is doing the thinking and the hearing and speaking all in one place, which means there's no translation and therefore loss of information that occurs when you translate from speech to text and from text back to speech. Um, so the current architecture is very limited actually, even though it's impressive, it's also very, very early. It's kind of like when chat GPT first came out, it was GPT 3.5 and GPT 3.5. If you went back in time and you used it now, you would find it to be a total joke. Um, and right now Grace seems really compelling, I think. I mean, I feel that she is hopefully our audience will as well. Uh, but six months from now you'll go back and say, Oh my gosh, that was such a simplistic thing. Um, so that's part of the reason to jump on it though, because, right, because it's not perfect and so will it be perfect in six months? Maybe not, but it's going to be a lot closer to perfect. Um, and so that's, that's the tech stack coming back to your question, Mallory. Does that make sense at a high level? Anything else you want to dig into?

[00:39:16:03 - 00:39:29:05]
Mallory
 It does. I want to know how you were able to give, provide grace with, uh, the knowledge base, like all of the knowledge of the sidecar site. Did you give her the link and say explore and go have at it? Did you like pre-program some stuff in there? What does that look like?

[00:39:29:05 - 00:41:25:17]
Amith
 So this is some of the software we built to make grace the system on top of the raw capabilities. So you have, you take the pieces and parts. It's kind of like, you know, you buy the engine for the car, you buy the transmission, you buy the wheels, but you have to like assemble it. And maybe you have to, you know, do some custom bodywork or something to make it look and work the way you want it to do. You know, the thing that you're hiring it to do. So grace, um, essentially is a system where we on the team built software that combines a bunch of things together. Part of it is knowledge. Um, grace is not nearly as knowledgeable about sidecar as the Betty instance of sidecar sub sidecar uses Betty, which is one of the companies in the blue Cyprus family Betty is this voluminous library of content. Like literally everything that sidecar has ever written or said, Betty's aware of. Grace only has a small subset of that. Um, because the technology behind Betty just takes longer to process a typical request to Betty. It's super deep and very rich in response, but it might take five or seven or even 10 seconds to come back to you. Kind of like when you're in thinking mode with Claude, it just takes a moment. That's not really suitable for audio yet. Um, those things are getting faster. So one day these things will converge where the best of the knowledge will come together, but what we did with grace is we built a very simple lookup system where grace knows a little bit about sidecar to begin with. And then grace has this continual knowledge of the conversation. Uh, and then grace has access to resources. So, um, those slides that you, that you saw that were displayed, um, those were essentially grace, um, knowing that there's 30 or so slides or what we call resources, um, that race can choose the display and each of those has a description and a title. So, um, that's the way it works. And then so grace as a system knows about these and based on the conversation, grace may choose to show you different slides. Now, when, when grace shows you a slide, she's not reading a prerecorded message.

[00:41:26:20 - 00:41:40:06]
Amith
 She's reasoned over what to tell you. Just like a good human salesperson might say, Hey, Malar, we've been talking for 15 minutes about this. And when I tell you about teams, I'm going to relate it back to, you know, the specific things that are important to you and your organization.

[00:41:40:06 - 00:41:59:15]
Mallory
 And I think you mentioned these slides or these images that grace has access to where those generated by AI, did the team build those? And then you mentioned they are titled. So she's making the decision based on what you say. She has access to this base of slides and then she chooses which one to show. Correct.

[00:41:59:15 - 00:43:55:05]
Amith
 That's correct. Yeah. And so the slides are pre-built. They're not generated dynamically. A lot of people have asked me that question and the way that works essentially is, um, you know, you have, let's, let's imagine you, Mallory were presenting to someone, but you didn't have like a prerecorded way of doing it and you just wanted to like show the pieces that made sense based on the conversation with the individual, but you know, there's 30 slides that you have access to and you have a list of them. So grace can pick the slide that she wants to show. Um, and that's part of what I was describing earlier where 11 labs has something called client side tools where the AI can say, Hey, I want you to display this image. And so then grace knows that that image has been displayed and then she can choose what to say. Uh, but those images themselves are predetermined. Um, the side cartoon builds, you know, 30 or 40 images using Google's nano banana two model. So the nano banana two is the new version of nano banana pro, which is both, uh, improved in terms of its texts capabilities, uh, even though nano banana pro from late last year was awesome with texts, uh, version two is even better and it's considerably faster and cheaper. Um, so that's what we used in order to build a library of slides. We do think that, uh, the next version of grace, which will probably be in the next 30 to 45 days, we'll have some dynamic content generation possibly. But another thing we're planning to do is to support video roles. So, uh, brief videos that will be anywhere from five to 30 second videos that will basically be like little mini demos. So someone says, Hey, what does it actually look like to be inside the side car LMS? How do I navigate around? What kinds of courses are available? Like, can you actually show it to me? Grace will be able to give you little pieces and narrate over the video. Won't have any audio built into it, but it will essentially be a video. And we'll describe the video to grace. And when grace asks for video, we'll play the video and tell grace the video is playing and then she'll start describing it. Um, so it's not quite like a live demo, but very close to that in some ways is how we think it will feel to people.

[00:43:55:05 - 00:44:10:15]
Mallory
 Wow. I was going to ask you about the roadmap for grace in the next, uh, a month or couple of months, how are you and the side car team measuring success for grace? Is it length of conversations? Is it conversions? What does that look like?

[00:44:11:15 - 00:45:52:09]
Amith
 That's an excellent question. We don't have specific metrics at the start. I mean, there's, there's several dimensions of this that I think are important. First, did we inspire folks to think about audio AI differently to me? If nothing else happened out of grace, other than we had even a handful of association leaders look at this and go, wow, I had no idea audio AI could do this. And I want to leverage this for my association sometime this year. That'd be a giant win. Um, of course you guys know I'm more ambitious than that. So I'd like to also see grace turn into, um, you know, something that's incredibly useful for people navigating their buying experiences, side car initially, and eventually you will see grace make her way into the side car learning hub itself. Um, grace will eventually be an active participant in the mastermind. So synchronous instructor led online learning where grace could be an active participant, um, grace will be essentially available everywhere that side car goes and be richer and richer, but that'll take time because the context you need to do that well is far greater. We kind of picked probably the easiest use case, um, for audio AI. Ultimately we'd love for grace to, you know, have a lot of successful conversions, conversion, meaning someone booking a meeting with our sales team, uh, or directly going ahead and proceeding with a purchase of the side car learning hub or signing up for an event. So we want grace to be helpful as a guide to guide people to the right things for them, if, if there's something that's a good fit, obviously. Um, and so from a commercial viewpoint, it'd be great if there were a lot of, of purchases, conversions going through that checkout process that you guys saw in the demo. Uh, and then finally, you know, we think this could become a product for blue Cypress as a family. It's, it's a pretty obvious one that this could be very helpful or it might become a feature of some of our existing products like Izzy or, or Betty.

[00:45:53:21 - 00:46:11:22]
Mallory
 Yeah. I think the opportunity here for associations is huge. Having something like this on your website or even, and I'm sure this is possible in me, right? We don't, I don't think we have a dedicated side car phone line. Maybe we do, but having a member call in on your phone number and having conversation there, I just feel like there's so much power here to better serve your members.

[00:46:13:08 - 00:46:57:17]
Amith
 I think the idea ultimately is, you know, and Grace said this well, right? It's, uh, meet your members where they're at. How do you find a way to deliver your unique value to each individual in your ecosystem, whether they're a member, a perspective member, maybe they're all one-time purchaser of a particular product or an attendee at an event. How do you meet them where they are at, um, in the modality that is most pleasing and most effective for them, um, and do it in a way that's deeply personalized, right? Grace not only was deeply personalized for us, but you know, she was not built to be a podcast host, but I thought she did a really good job of jumping on the pod and, and immediately adapting to the context of that conversation.

[00:46:57:17 - 00:47:35:13]
Mallory
 She did a great job. And as context for the listener, before we recorded this, I prerecorded a demo and Amith had the idea, well, why don't we have her join the podcast? And I was thinking, I don't know if she can do that. Like she's built to talk about sidecar and sure enough, I was even surprised with the black box that is AI. So I feel like she did a great job. Grace is one of the more exciting things I've tested out personally in the last few months, seeing that on the sidecar site, I immediately messaged Amith and was like, she is incredible. Uh, of course we're going to keep you all up to speed on that. Amith, any final takeaways, parting thoughts on this audio agent themed episode for our association listeners?

[00:47:35:13 - 00:47:53:10]
Amith
 Uh, nothing really. I'm excited to hear your reaction though, that it's one of the more exciting things you tested in the last few months is, you know, part of what you just enjoy doing and what you do for sidecar is to stay aware of all the latest and greatest in AI, so, and there's a lot happening in the world of AI, even just this year. It's, you know, we're not even three months in and there's a lot of crazy things that have happened.

[00:47:53:10 - 00:48:17:18]
Mallory
 A hundred percent. So if you all are still with us, go to the sidecar site, if you haven't already, try out grace for yourself. If you're listening to this episode audio only, I also highly encourage you to check out the YouTube version, seeing those images pulled up on the fly as someone, someone has an AI is walking you through all of your AI education options is just incredible. And I think it will be quite inspirational for you. If you have any

[00:48:17:18 - 00:48:23:05]
 (Music Playing)

[00:48:33:23 - 00:48:50:22]
Mallory
 Thanks for tuning into the Sidecar Sync podcast. If you want to dive deeper into anything mentioned in this episode, please check out the links in our show notes. And if you're looking for more in-depth AI education for you, your entire team, or your members, head to sidecar.ai.

[00:48:50:22 - 00:48:54:03]
 (Music Playing)