How Caritas Turns Meetings Into AI-Ready Context (Full Transcript)

[00:00:00] Speaker 1: and we are live. I already see some folks joining. Hey Hugo, Rhonda, welcome everyone. I have Christian here with me today. Welcome Ryan. Just gonna give a minute for all members to join and we can get started without delays. Hugo is joining from Hudson, Ohio. We always ask because we have such a diverse and international audience. So it's very curious to see where everyone's joining from. And I see a lot of folks joining. Scott, Matt, Jeff, Emmanuel, welcome everyone. Daniel, Bronwyn, Mother Scott, Jacob. Yeah, so welcome and welcome. Thanks so much for being here, everyone. This is our first very new session for Firefly's Customer Show and Tells. I'm really excited to start this new format here. Instead of us showing the features, we invite a customer to show how they actually run their work, business, and tasks through Firefly's. And today that's Christian, founder and principal of Caritas Venture. And before we start, again, to drop in the chat where you're joining from, what you do. I'd love to know everyone that's here in the room. From Canada, Hugo's asking where you're joining from. So Christian's joined from Philly, that's correct? That's right. And a quick word about Christian. As said, he runs Caritas Venture where they help business owners put AI to work and build capabilities that actually help. And Christian has an MBA from MIT Sloan and was the co-chair of the MIT AI Conference at Caritas. Firefly's is a core source of context for both client work and international operations. And that's exactly what we're going to see today. But Christian, if you would introduce yourself better and let everyone know more about

[00:02:24] Speaker 2: you. Excellent, Pedro. Thank you so much. In addition to being a Firefly's advocate and super user, it's been core to our business operations for about four years now. I lead a team of consultants globally distributed, North and South America, and then in Europe. And we work together to build software where there is none for our clients and upskill the operators who maybe have no technical acumen so that they can build and maintain that software without us, become more productive and make the companies more valuable. So we've been doing this for about four years and Firefly's has been absolutely essential to making this work across the board. And I'm excited to share some examples today. So glad to hear. And so how this works?

[00:03:14] Speaker 1: Christian will actually share the screen. Christian will share his experience and walk us through real workflows and examples from Caritas. You can drop your questions at any time in the chat. But we also have a Q&A feature is the question mark icon here at the bottom in the middle. I'll be collecting them and bring them to Christian at the end. So don't hold back. Drop at any time you have a question or want to know more about and we are recording. So we're going to post everything in our resource libraries after the session. And you're going to get also a recap in your email. And today we're going to cover what most things treat meetings as something that happens then gets better. But different from that at Caritas, every conversation becomes shared context for the whole business. Time calls and team meetings all end in Firefly's. Then another AI tools pulls all that context to build, brief, and draft across the business. The people stay in the loop this way. The team, often the customer as well, and reviews work before it goes anywhere and gets visited. And as you watch everything that Christian is going to present, be sure to think about which conversations you have during your week, during your day, and how you want to put that to work. Similar to what you're seeing on screen. And while you think about that, those are the best questions for the Q&A. Do drop anything that you see here live and want to have for yourself as well on how to apply it. And last thing before we pass it over to Christian, for anyone who's new to Firefly, here's a quick picture on everything we enable Firefly. Now it is your AI teammate across your whole workday, not just meetings. It captures your conversations rather than just meetings, emails, lag, calls, integrations, and keeps everything organized in one place with your own company knowledge base. You have tasks, you have daily briefs, and puts all that context to work with ASQs, integrations, and voice agents. This one here for today, the MCP is what makes Christian's workflow possible. It lets AI tools like Cloud pull all your meeting contexts directly and you'll see in action how that works in a minute. And for anyone who also don't know, Firefly is used by 20 million people across the world and 75% of the Fortune 500 companies. All right. So that's enough from me and Christian, over to you. I will move you to your own slides.

[00:06:15] Speaker 2: Excellent, Pedro. Thank you so much. So we have five different use cases that we're going to dig into today. If folks have questions, I'm just going to kind of go through these sequentially, but definitely put them in the chat and we'll get to them at the very end. So we'll try to do 15 minutes here and then 15 minutes for questions. And these use cases are going to cover everything from core operations, how we actually serve our clients, to keeping the team up to date, keeping sponsors sort of in the loop on what's going on, and trying to make sure that our overhead and running the business is sort of low enough that we can handle all these complex different projects consultants are handling at any given time. One to five different fully custom builds across different industries and Firefly being this glue or AI being the glue that allows us to pull together context across everybody's meetings, get folks up to speed, and execute well has been a core part of our business from the beginning. So first one I'm going to go into is live development. So we click in there. I'm going to talk through this. All right. So I like to think we've been vibe coding since before vibe coding was a thing. One of the early insights I had going back all the way to GPT-4 when I was begging Logan who was at OpenAI at the time and DMs on Twitter to give me access to the GPT-4 32k context window model was that if we asked really good questions during meetings, then we could get sufficient context that actually building at the time, you know, simple scripts, but then eventually web applications and now much more sophisticated software was a lot easier because what we would do is we would, you know, do the discovery. Sometimes it's one, two, three hours of stakeholder interviews. Back in the day, you can only fit, you know, one or two transcripts into a single prompt. And now it's totally flawless because of the MCP connectors and whatnot. But if you have those recordings, you can then prompt Claude to read them in their entirety or chat GPT or whatever model you want to use, and then essentially turn that into software, right? There are a few steps along the way. You got to guide the models and everything, but actually collecting that context in the form of a transcript, asking good questions to the person who might use it or whoever the subject matter expert is, has been core to our approach from the beginning. And one thing that's amazing is Firefly's added a live MCP tool. So MCP is the model context protocol. It's sort of like, you know, apps for agents, right? So if you have a connector to Claude or chat GPT, it's using an MCP interface on a generally a remote hosted server in the background. There's an MCP tool that allows you to pull in the transcript as it's being generated of the conversation that you're having. And we have our settings in Firefly set up so that anybody who's a meeting attendee also has access to the meeting that we're recording, which means when we work with clients, we will actually have them set up Claude code and the various bits of tooling that we recommend on their local machines. We will have them log into Firefly's and add the MCP server. And then not only can you use the transcripts after they've been recorded to help build software, but we will actually do this live. So as you can see, I'm outside. I'm a big advocate from, you know, work from anywhere. And I like to get my 20,000 steps in a day. And we've now gotten to the point with clients where I and my collaborators are able to have a meeting. And in this particular case, I'll share with you, I didn't even look at the web application that was being built as a part of the engagement that we had. But based on the feedback of folks during the call and sort of as a function of asking the right questions, the person who we were working with, with that live connector was able to have Claude code pull continuously to sort of listen to the live transcript and iterate and make changes as we were having the discussion. That was a really powerful thing. And I'll show you and, you know, to that person's satisfaction. So the feedback loop on this kind of thing is incredibly tight, and it wouldn't be possible if it were not for Firefly's live. So if you click on the, click on that first one, you can actually see this is a redacted screenshot from one of our sessions. My camera was off because I was literally walking on this trail that I'm on right now. But basically, we had one of my collaborators, one of my consultants, and then the client. And, you know, this was, I guess, like an hour-long call that we had. And there was the terminal open, and then the web application running locally on his system, powered by data that he had connections to. And just simply through, you know, asking the right questions, saying sort of the right set of magic words, using the right terms of art and whatnot, we were able to produce something that ultimately has massively boosted the productivity of this particular client. I think there's maybe like one more screenshot. I can show you where you see the terminal running. Yeah, so you can see here, I mean, he literally had the terminal open with Cloud Code, and we had prompted it at the beginning, hey, you've got access to the Firefly's MCP live connector, just listen to it, and decide to make changes, like as you see fit. Right, if you want, it depends on the degrees of autonomy you want to give it, but that was how that worked. All right, so that's first one. We'll go to the next one. All right, staying close to, oh, okay, so my team is asynchronous. They're also all fractional, meaning they're doing these, long story short, is the way that we do client engagements is everything is bounty based. So my consultants, they have skin in the game, they make a lot of money, and can in a very short period of time if they're successful, and they achieve the outcomes we negotiate with clients. But they're responsible for achieving those outcomes. But they're responsible for achieving them, and so they share in the downside as well. And so in line with this culture and ethos of giving people maximal degrees of freedom in terms of how they do their work, you know, there's still the sort of matter of ensuring quality, especially for newcomers, because we have a very liquid team. It sort of goes up and down, depending on people's bandwidth. And so what I built was a system that plugs into Firefly's programmatically through I think what's called the Super Admin API. So we have an enterprise account, which allows me to programmatically query transcripts across the entire team. And I built a tool called Powerline, and it essentially pings me on Slack after every meeting, either meetings I was in, or meetings that I was not in, that the associates were in, and it gives me kind of a chief of staff style read. So instead of me having to get caught up at the end of the day, or only catch up during one-on-ones, I get these little bite-sized bits of feedback. And sometimes I will forward that Slack message, you know, in a DM, or sometimes in a group channel for the client, you know, tag the person who's in there, ask questions, occasionally give coaching, or, you know, look at like proactively unblocking them, for example. And so being able to have these asynchronous pings has been incredibly valuable in keeping the team sort of on the same page, and in a way that respects everybody's time. So if you click on the first example, yeah, so this was like, this is a highly redacted one, that's how you know it's good information. So this is like a screenshot of one of the meetings I was not in that I get pinged with privately. And, you know, it's just like a little three-paragraph thing. It gives me like, in the prompt, and like give me subtextual reads, and, you know, maybe like unspoken dynamics that I should know about, or whatever the case is. And it does a good job, because it has access to the full transcript. And then I decide if there's anything I need to do, or if I'm just in the loop. So that's like one example of a meeting I wasn't in. And then that's one that I was in, and this is called a hot wash. So, you know, what's something that I maybe wasn't paying attention to that's important? I just get a little ping at the end, and it makes me think a little differently. And then it gives me a quick win opportunity. So like, what's one action I wouldn't have taken? Sometimes it's, you know, text the sponsor, or reach out to this person, or, you know, just do something small. And so it's like all these little, these little bits of feedback have made it a lot easier for me to keep momentum going and make good decisions maybe I wouldn't have otherwise. And then I get a weekly email, which I always look forward to. Actually, the first time I started setting this up, I was kind of shocked at how good it was. But I feed the system all of the Firefly's transcripts across everything that the team has done over a week. And I ask AI to, in this case, we're using Gemini for this one, but I get an email, like I would from a trusted advisor, bringing things to my attention and synthesizing themes across all the different transcripts. And I found it really useful. It makes it easier in a lot of ways for me to shut off at the end of the week, because I get this nice bookend that kind of ties everything together. All right, so that's Powerline. And then this, okay, so let's talk about onboarding. So as I mentioned, the team is like very liquid. And again, I try to have us have as few unnecessary meetings as possible. Ironically, you know, we're all spending more time in meetings, but it's, but it's, it's with clients, right? Like that's a good use of attention. And so traditionally, it was very expensive to onboard people, you know, to a new industry that they have no experience in, people they don't know, like there's so much context. And in our work, everything is super high context, right? Like all the details really matter. And Firefly's has been very helpful in this regard, in that as soon as somebody comes on board, if they're being onboarded to a new project, I will kick off agent threads. Or if one of our senior consultants is onboarding a junior consultant, they'll kick off an agent thread that will both take all the salient Firefly's transcripts, let's say for a particular client engagement, send it to them on Slack or email or whatever the right channel is. But then, and this has been critical as a Firefly's feature, share with them the actual access to the transcripts in a way that's, that has an expiry date, just, just in the interest of maintaining, you know, info security best practices. So it's, it's, you know, it's not shared in perpetuity. It's, there's like a countdown clock on it. So has made it very easy for people to get up to speed. If they have follow up questions, then they can go query the primary sources directly or watch the tape. And it's massively reduced the time it takes to become productive for new members. So it's been great. And yeah, and that's, that's an example. I think in this way, this one, there may be, usually like these Slack messages will contain links to the primary sources as well. So the Firefly's meetings, and, and again, MCP is really useful. It's got that share feature with the expiry parameter on it. All right. So, you know, a big part of our ethos when it comes to marketing is creating and sharing compelling facts is the way I put it, right. And it's always been a best practice to create case studies, but what's awesome about recording everything with Firefly's is you have the entire story there, basically an excruciating detail. And as, again, as the models have gotten more powerful, it's gotten easier to do this. At the end of every engagement, we have a formal internal reflection. And, and we, we drafted an internal private case study that we keep in our repository that people can refer back to, you know, in case they have a new client and they want to see what have we done in the past that might be similar. And that's distilled and derived from Firefly's transcripts. And in a lot of cases, the claims are then footnoted to the primary sources. So an agent can actually traverse a graph and go see, see what happened starting at the case study level. And then externally, if you go on our website, we've got, I think now four or five case studies that, that clients have signed off on. And what's, what's amazing about, yeah, there's one IGS, great. They've been a fantastic partner. What's great about this, I'll give you a very concrete example for IGS. So we ended up, we did one case study, and then we wanted to do some updates because I had a request from somebody to send an updated version of that case study to somebody else who was at a different company. And so all I had to do now that we have this really powerful agents was to spin up, I think in this case, it was a five points, Codex 5.6 Sol Ultra, some kind of state-of-the-art token intensive thing where I said, okay, go take a look at that request. I want you to now draft a second case study. And it would do that, we're on Netlify, so it'll produce like a PR with a Netlify preview. And then I want you to take all that, draft it and, and then send it, send it over to, to the executive sponsor to get their eyes on it and get feedback and then iterate while having access to all the primary sources until the preview is to their satisfaction, thus completely taking me out of the loop and getting it until, you know, she had signed off, one or two other stakeholders had signed off, that was all in email. And then it was able to publish it to production and then send it to the person who had requested it. First of all, it's incredible that technology in general is that good. But what was key to making that happen was the fact that it had the MCP-mediated access to the transcripts where it could fact-check things, it could, you know, retrieve information, it could do the updates, and I could be essentially completely out of the loop for that. So that's, that's an example, and then it's on the website and it's live. So that was pretty cool. And of course, if you're curious about what we do, I would welcome you to check out the case studies on the website. We'll go to the next one. All right, I think this might be the last one, but we run on an operating system called Arlos. This is like a custom proprietary bit of technology that we developed to support our outcome-based engagements. And one of the ways we keep everybody on the same page, you know, as a good Bayesian does, is to update probability scores of, okay, given everything we know, this is what we think the odds of success are, given that we have X number of weeks left, and so on and so forth. And so this is super useful because those probability scores are about as signal-dense as you can get without something actually being, you know, priced in a market or something like that. And if I see, or if a sponsor sees, or whoever gets these automated email updates when probability scores get published, they can see that something went from like 40% to 70%. That's kind of an invitation to dig in and figure out, okay, like what changed and how do we replicate that? Or if something dropped from 80% to 20%, then it's a question of, okay, what can I do, you know, at a higher level of management to intervene, to unblock folks, to use whatever was learned in that context to help boost the odds of success. And in terms of the person from our side who's responsible for achieving this outcome, in the Arlo system, they're called the driver, they actually are able to draft the probability scores in the first place, and then they apply their judgment. They'll, you know, give a final score by having Cloud Code or Codex or whatever client they want to use to go trawl through all of the transcripts, let's say from the past week, and then look at that versus where everything was the week before, and then where we're trying to get to and help draft that score and give a little bit of reasoning. So having MCP mediated access again to Fireflies has been completely essential to make this happen. And I think that might be it. Let's go into questions. Also, if you want to connect with me on LinkedIn, I respond to every message I get. And if you want to check out our website, it's at karytas.ventures. So definitely connect with us afterward as well. But I think we can

[00:23:30] Speaker 1: dive in and do some questions. Definitely. And thank you, thanks so much, Christian, that was amazing to see. And we do have some questions here. We'll start first from Daniel. And saying that Firefly is great for capturing words, but a very important part of conversations is context. So assumptions about who's who, the power dynamics, and all that relates to that. And in your view, by doing all these cases, do you think that by us relying so much on the AI solutions, we might miss what's actually going on? Or it actually helps getting all the contacts in the same place, and connecting? Because when we are in a meeting, we are present in the meeting. But if we miss it, and any other related or past one, we don't have the full context, do you think that AI either helps with that generally? Or it still misses some parts of it?

[00:24:44] Speaker 2: Yeah, I'd say it's on that very helpful. You know, what's great about MCP in particular, now that it's easier for the agents to pull context in one fell swoop from a lot of different sources, is you can take into account way more high fidelity information than you can hold in your head. So if you're going into, I mean, I've had this happen before, we're really high stakes negotiation. And I will, as part of the preparatory process, have an agent go pull all the transcripts of all the interactions that all my team members have had with the other team, including, you know, calls going back weeks or months or whatever the case is, and then drive the agent to help me pull out subtext, in the interest of maybe bringing things to my attention that I just didn't notice or was unlikely to have noticed. And so, and so I would say on net, it's useful in that context, or it's useful towards this end. It's actually, you know, if you want to freak yourself out a little bit, you know, a prompt that you can put in if you have Firefly's transcripts say, you know, I'm almost hesitant to say this, because I know, like, anyway, I'm just going to say it. But there's, there's a prompt that you can give saying, hey, I want you to look at all the Firefly's transcripts of the interactions that I've had with this particular person, like whatever it is. And say, I want you to give me a deeply Straussian read. Leo Strauss was this 20th century philosopher who believed that, you know, all philosophical texts had like hidden meanings, basically, right. And so it lends itself a little bit to like, paranoid conspiratorial thinking a bit, but for sure, it'll bring things to your attention when it comes to social interactions, power dynamics, those things left unsaid, that may be lurking in the background. Now, the caveat to that is this is just a lens that you're putting over data, right? So like, it's like a microscope that's maybe allowing you to see things that were hard to see before. But you as the person making that decision based on that have full responsibility, and you need to apply your own judgment and course correct and not completely rely on the model. So it's a different lens, you can put over all this context, it absolutely can elucidate some of these more subtle, nuanced interactions, but whether or not it's also true, and if it's useful to think of things in that way, that's a choice that you have to make. But yeah, in my experience,

[00:27:09] Speaker 1: it definitely works. It was a good question. Definitely. Yeah, I think like the best scenario for AI currently, it is getting rid of all those minor tasks and those daily workflows that we have to do and just keep the human on the review parts before anything goes live or any agreements are made. But I also want to hear from everyone here, what parts of your workflows you keep a human in the loop? Or is there anything that you let AI do 100% of it? Let me know in the chat as well. And also, Daniel, it is also appreciating you, Christian,

[00:27:55] Speaker 2: for introducing Kefir Appliance. They are loving it. I've been an evangelist for a long time, certainly. And Daniel, good to hear from you. So the question was when to stay in and out of the loop. That's a really important, big question. I don't know, a couple of things. I guess like one is you can't actually take the human out of the loop. It's just like, it's impossible. You need somebody making judgment calls. You can absolutely expand the size of your loops, so to speak, and you know, in some sense exacerbated by the fact that the technology is constantly changing and becoming more capable in a lot of ways. I would say that this has remained very much an art rather than a science. The science part of it has always been basically you need two ingredients to do well with AI. One is a good goal. So having a good sense of what it is you're trying to achieve. If you can specify that and formalize it in plain language, you're going to be able to make this stuff more useful. And then two is context, right? And Firefly is, you can think of as like, you know, sensors as a service or something like that. It's, you know, continuously measuring in high fidelity what's going on with your business and these really rich, complex interactions that are happening internally or, you know, externally with clients. That context is absolutely critical to having the confidence to give the agents more autonomy, because you can be confident that what they're doing is at least informed or somehow fueled by ground truth, which exists in the transcripts. Like we can all basically agree that if somebody said something, they may have literally said it, there's a lot of more context, there's, you know, nonverbal cues and everything that are important that maybe don't show up in the transcripts per se. But as far as keeping these things grounded, transcripts are, you know, hard to beat as far as a type of data source. And so as the models get better, and you want to expand the size of your loops and take those risks, just like I mentioned with the case study example, right? I had a good, highly specified kind of verification mechanism, which is that the sponsor would actually sign off and, you know, give formal approval in written form. You know, I trusted enough to iterate sort of on its own without bothering me about the content. And that subsequently has worked. So, yeah.

[00:30:26] Speaker 1: I like it about something that you mentioned while you were presenting that the also the best part of having these workflows is that knowing by what you also expect from the output of the workflow, you also have better meetings because you remember that things will get captured. And then you remember that during the meeting to be sure that you either discuss, ask, or point out, voice out that during the meeting. So, it also gets captured later as well. And Sam also asked something very nice here that they're not sure the value of having the access to the transcript live via the MCP. And what kind of changes are you making live

[00:31:16] Speaker 2: or is just to be more accurate? Well, it's you're reducing the iteration or you're compressing the iteration loop. So, you know, this basically takes two forms. So, thinking back to that example of those redacted screenshots that I shared on that very call, a couple of things happened. So, one was I sort of like slowed down at some point and I said, hey, so and so, like, can you just take a second, look over the dashboard that was created. And, you know, this person's very, you know, it's a very high context task where they've been very close to the data in the process. So, they kind of they have a good gut sense of whether something looks right or wrong. And I essentially said, like, does anything look off? Does anything look weird? You know, what are any questions that just pop into your head? Just create some space and, you know, verbalize it, right? Like anything that doesn't make sense. And let him say all those things. And because Claude was polling live, like, we give it a prompt at the beginning to just listen to what's going on and then react to it accordingly. He got it all out and then Claude chugged away and answered those questions. And I think in one case, actually, because he called out, you know, something was some way about the way the data was represented was not right. It actually went ahead and actually fixed the web application in real time. And so, asking questions to get clarity and making sure that you're able to understand sort of what's going on, the logic of the product or whatever else, you just wouldn't have been able to do that in real time before. And you would have had to do a lot more clicking or maybe copying and pasting and things like that. Now you don't have to do that with an MCP server and an MCP client, just pulling it continuously, that live tool, live transcript tool. And then two is giving a direction. So saying, well, I want more of this and more of that. And, you know, it'll just go ahead and do it. And while it's chugging away, I don't know, we can talk about other things and get to know each other. And then once it has something for us to react to, we can just continue going. So it's massively reduced the friction and also compressed the cycle times for iteration when doing product development in this like highly collaborative, high context way.

[00:33:42] Speaker 1: Which is actually also why we have with the Fireflies either Chrome extension that you use or the desktop app, we have the transcript live and also the live assist where you can keep asking things while the meeting is happening. So it's a similar use case to that. And Joan is asking that since you mentioned the concept of updating and you are dealing with clients over time and your own team over time, have you used the internal reflection within the conversation of the device updating from AI as how AI perceives the personality of the client or employer? We as humans change over time. I'm curious if AI is also recognizing that and capturing over your workflows.

[00:34:33] Speaker 2: So I'm trying to understand this. Is the question basically, are the systems able to take into account the changes that we're making in response to the like the signals that we're using to make those smaller changes?

[00:34:50] Speaker 1: I think it's related to that more on possibly how it captures not only let's say the project context and how the, let's say the deal or any company it finds is evolving, but also the personality

[00:35:13] Speaker 2: of the source that is coming. That's a great question. I think it's yes with a caveat. So when it comes to updating or learning in some sense, at the end of the day, the way these models work, and this is kind of a technical limitation, at best they're doing what's called in context learning. So they have some kind of very general worldview that's baked into the values of the weights during pre-training, post-training, everything else. And this worldview is somewhat static. I mean, the model's being updated periodically, but they're not really, they're not being updated on your private data. So what's essentially happening is if we, if you were to ask the question of, you know, given sort of all the Firefly's transcripts from the past three months, you know, how has so-and-so evolved and changed? It would be able to give you an answer because it's going to line up all those full transcripts like in the context window and then try to compress it into something that will like turn into words and give you an interpretation. But it's not updating the weights itself. So it's not sort of learning in the sense that we are, it's more doing this like pseudo learning compression activity that is maybe useful for teasing out some of those trends. But it's maybe not as nuanced, not as nuanced, you know, not as reflective or attentive maybe as a person would be with the

[00:36:51] Speaker 1: same question. Actually, we also do a lot with HR and recruitment for Firefly's cases. So we even have some AI skills that relates to that, that a lot of folks use to get more of their personality. And Ilia sent a very great question that we get all the time. And I will even add a note to the question. But don't you think in a world where AI tools can record and analyze your meetings, negotiations, work in general, professionals might lose focus during these meetings? Or these are my own notes. Do you think that we actually have better meetings or even less meetings because we have enough context to avoid having to go live and just ask, hey, what's the update?

[00:37:48] Speaker 2: Yeah, this is a question from Ilia.

[00:37:54] Speaker 1: Is that right? Yeah. Okay. I just added some notes, but going back to her question, do you think that with all this AI workflows currently, professionals may lose focus during the meetings to kind of maybe assuming that they will get the summary later?

[00:38:15] Speaker 2: It is a good question. I've heard that feedback and I'm sure that I myself have been less attentive because I know in the back of my head it's going to be recorded. I'm not sure if that's necessarily the case. But I think it's a good question that, you know, this is a very interesting question around these meeting recorders and I think social norms in general, because the other challenge that can happen is, you know, there are conversations that are going to be chilled, like there's a bit of a chilling effect, right, if people know that they're on the record. And so I think, you know, this is all being figured out in real time, but I would say, you know, if you want to create the conditions for people to, let's say, be more attentive and speak more freely, sometimes you got to make the call. And, you know, I'll do this with clients occasionally, where you're just going to use another channel and everybody understands that it's off the record, in which case, you know, you may be more attentive and you may be able to speak a little bit more freely, but you don't get the advantages of the AI analysis. So I think it's a really good question. I think more generally, we're going to have to navigate this overall change in norms around this stuff, because we are a new territory. But I would say, too, if you have the note taker there, you can still be attentive and probably should if you want to produce the best data that's going to be useful in the context of AI.

[00:39:54] Speaker 1: I agree with that. Also, Arthur sent another great question, because I mentioned the live assists, and it also relates to your workflow. And for people that are getting started, should they start with our own live assists and use our live transcription or move towards more a solution like yours using the MCP with more customized outputs and workflows?

[00:40:30] Speaker 2: I would say whatever is going to be the least amount of friction. You know, the live assist tool, the live assist feature is fantastic. You know, the Fred sort of pain to be able to query it through that interface if you're comfortable working in that way. And I think one of the things that's awesome about Fireflies and why we've we've loved working with them, and it is a core part of our system, is they make it really easy if you want to be a power user, and you know, combine the data with other sources in the way that you want to get it out with the power tools. And so, and you know, the programmatic connectors through MCP and the API and everything else. So yeah, I would say it's a great product, push it as far as you can. And if you need to go for something a little extra, fortunately Fireflies is one of, you know, I think the only vendors in the space that really makes it easy for you to take things to the next level if you want to do that, so.

[00:41:25] Speaker 1: Also very glad to hear that. And I think we have one more question here, but I think we have time for another one. But just because Hugo is standing here, Christian, that it's great to see a fellow Blue Devil and asking about your experience at Duke. But he is asking other AI tools that you're using, specifically also pocket recorders, that me and Christian was actually talking very before

[00:41:55] Speaker 2: this very meeting. Yes, it's a great question. I always get to connect with another Blue Devil, go Blue Devils. Yeah, so I graduated from a Trinity econ degree in 2013. Awesome school, some of my best friends I made there. And I would say many, many good memories, both on the academic side and then also socially. Greek life at Duke was a lot of fun. I don't know, it's hard to not think about tailgate. If you went to Duke, you know what tailgate is. You know, it was sort of a rather raucous, fun compliment to the football games. So I think going to my very first tailgate was a fun and maybe kind of shocking experience the first time. But that is something that comes to mind. In terms of other AI tools, that's a good question. So dictation, which I think Fireflies now has its own dictation tool, I think is absolutely essential. So doing something where you don't have to type anymore, you can do it on iOS, you can do it on your desktop. So absolutely dictation, stop typing, start talking is advice that I give to clients, usually on day one. And then in terms of the models themselves, I mean, Anthropic and OpenAI definitely have like the best foundation models, that just kind of is what it is. So Cloud Code either through the terminal or, you know, through the native application is really good. Same thing with ChatGPTN codecs. If you want to get a little bit more customized, I'm a big fan of Hermes and this company Noose Research, where if you want to set up like a harness that's a little bit more extensible, and either run inference through hardware that you control or through other, you know, model providers through like an open router or something like that, Hermes can be pretty great. And then one thing I'm a huge fan of is using these remote control features on the native application. So you can do this through ChatGPT, through the Cloud application, you can actually connect to your computer where everything is running back home. So your laptop and a lot of people bought Mac minis and things like this for this purpose. And that's been essential to me changing how I work where I get to be outside getting 20,000 steps a day, you know, in beautiful suburban Philadelphia. And either I'm taking calls to get recorded through Fireflies, or, you know, I'm prompting the agents to go do things, coordinate with the team. I've got one running in the background right now who's onboarding somebody as we speak, a new associate. Having the remote control feature set up has been essential to, frankly, making work much more of a joy. So I would add that as well.

[00:45:02] Speaker 1: And one last question here from Jacob, which is also a good one. We kind of discussed this, but do you review each of the summaries of each meeting and amend any issues? They are having some issues with the speaker names being correct, especially a problem with the external contractor when they have no contacts already. Like who should be doing something particular? And their names start appearing in the action items. It's a great question. I do, we have that custom

[00:45:38] Speaker 2: power line system that we use, which is kind of similar. But that, I mean, yeah, that's just the answer of what we use as far as like the summaries. And, you know, condensing the meeting transcripts into something that's going to be more useful in the context of the business. But the Firefly summaries are good. In terms of amending issues, it's a good question. I, one of the reasons that we chose Firefly, or I chose Firefly as the primary tool was because speaker identification was so good. So, yeah, it's a good question. I've not personally really run into that issue. As long as you have the note taker join the meeting, in which case, I think it's, I don't know what's going on behind the scenes, but it does a very, very, very good job if it's actually in the meeting, generally speaking, of diarizing the speakers. That may not be the case with like a Chrome extension or something. But certainly if you have the assistant join the meeting, which is what we do by default, it tends to work well.

[00:46:41] Speaker 1: Glad to hear that as well. And if you have not tried this after today, start by connecting Firefly to the AI tool you already use, like Christian is doing with the MCP, and save the date for our next events. Next week, we have the show and tell on Tuesday, where we're going to talk about coaching discovery calls. Christian, thank you so, so much for sharing all this with us. If you want to learn more about Caritas, head to caritas.ventures. I shared it in the chat. Thanks, everyone, again, for joining. The report will be in the resource library. And before I go, leave a comment to Christian here, and feel free to connect with him in the community. Go there and search Christian. He will be there. Perfect. Pedro, thank you so much. It was a pleasure. Thank you. See you. See you, everyone in the community. Until the next one. Bye.