TezTalks Radio - Tezos Ecosystem Podcast
TezTalks Radio - Tezos Ecosystem Podcast
126: TezTalks DeFi | Kitsune from Keystone Labs
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
We talk with Kitsune from Keystone Lab about Panora, a mobile app that captures real world conversations and turns them into transcripts you can search and chat with later. We dig into the product decisions behind mobile first capture, privacy focused processing, and how AI helped ship an MVP fast while distribution stays the hardest challenge.
• why unscheduled conversations carry the most valuable work context
• who Panora is for, from working professionals to lawyers, HR, and social workers
• the mobile first workflow, live transcript, and post recording processing
• on device transcription vs backend processing, with encryption and user control
• artifacts like topics, events, and decisions, plus chat based recall
• optional voice tagging to group recordings by person
• long term memory via indexing, recency, and lightweight semantic search
• how AI accelerates development, with multi model reviews and harnesses
• lessons learned, especially scope creep and shipping tradeoffs
Welcome And What Keystone Builds
SPEAKER_00Hi everyone, welcome to a new episode of Test Talks Radio. Today we have Kitsune from TZAP coming here to share with us about Keystone and Panora. Kitsune, kindly sh uh introduce yourself and briefly about uh Keystone and Panora.
SPEAKER_02Hi Anthony. I'm uh Kitsune. I'm a junior developer at TCAPEC. TZAPEC, sorry.
SPEAKER_01Uh how do I say this?
SPEAKER_02So TCAP was running accelerators and incubators for the past few years, with 45 laps being the latest iteration. But this year we are focusing on internal product development, guided by everything we have learned over the past years, and we're doing it through Keystone Lab. And the first product that we have completed, at least at its current stage, right, is uh Panora, which is uh meeting notes. Well, not exactly meeting notes, but any kind of conversations. As long as conversations are crucial to your to your job or to your workflow, Panora will be there to help you contextualize and organize the information that you are faced with.
SPEAKER_00Got it. And yeah, what inspired you guys to have uh to build Panora as your first product? And yeah, maybe later on you can also share a bit on uh what are some key lessons that you guys uh brought from uh incubator days uh at TZAP to uh Keystone Lab.
SPEAKER_02Yeah. So the original implementation is a simple observation.
Why Unscheduled Talks Matter Most
SPEAKER_02The conversations that shape our work are often not the ones in the calendar. Like right now we have many, we have multiple tools to capture scheduled online meetings, but a lot of the valuable context happens outside the meeting, sometimes like during uh a side check with the decision makers or in a conference, or maybe just by the water cooler, I mean if your office has that. And these conversations some often contain the things that matter, like it could be the context on why decisions were made, or what someone promised, and what those concerns raised, but because it's not captured within the meeting itself, this information disappears. So you saw this gap, and this was happening multiple times, especially when everything is quite uh agile in your decisions when we are trying to work with the founders to work on different strategies to improve their product or improve their distribution or their reach. So a lot of these decisions were made on one-on-ones or basically outside of the conventional meeting areas, and we feel that we felt that this this could be captured which or should be captured better instead of just relying on memory. So Panora is more for the impromptu conversations that happens, like, or any conversation that happens for that matter. And yeah.
SPEAKER_00Nice. Uh yeah, that's that's a very interesting angle. Yeah, I I do notice now Google Meets uh most yeah, has Gemini to cover uh summaries, but yeah, in person, that's a huge gap that hasn't really been solved.
Mobile First Capture And Workflow
SPEAKER_00For Penora, you guys are predominantly a mobile app, right?
SPEAKER_02Yes. So I mean we have gone through back and forth about the architecture of how they should be. Like I recognize that work should be done or is more productive to be done on the computer, but if you have the computer, you are more likely to have access to much more tools. Whereas when you are just having a side chat with your colleague or with your managers, it's mostly one-on-one, and the only device that you have with you will probably be your phone. And the phone seems less intrusive in a way that it's it's it's just a small, a smaller device just for the meeting capture. Then for now, in the current iteration, most of the recall and memory work will be done on the device. But in future, if we were to progress, right, it'd be on the roadmap to have this as a as a web app where you can really dive in and do serious work after the meeting. So it remains mobile first because the capture happens on mobile, which is a more convenient way than to log your laptop around, right? Yeah, but after that, when the work really happens, you can either export it to your computer or in future you will have a we might have a web version of this to allow you to work on it.
SPEAKER_00Okay. And I guess based on the examples you've given so far, the IB user profile is uh a working professional, right? Or or does it just is it broadly everyone that you see uh being used for?
SPEAKER_02Yeah, that's a good point. It's not just I mean it can be any working any kind of working professional that relies on conversations, right? Uh it could be lawyers, it could be uh HR person interviewing interviewing uh like new new hires. Or it could be social workers, as long as the context and the gist of the work happens through conversations, it can be a useful tool to capture that.
SPEAKER_00Yeah, I agree. Okay, okay. Um yeah, and for the for that, let's say let's say I'm using the app right now. What's the typical user journey? I I I'm maybe having a conversation, I press record. Uh yeah, maybe you can walk me through uh how how that whole process goes. Okay.
Privacy Options And Encrypted Processing
SPEAKER_02So right now I'm also recording this uh this conversation between you and me. So what we do is we open the app, we press record, and when it's recording, we will be able to see as a live a live transcript of the stuff of the words that we have spoken. And we can also see like how many speakers the app thinks that there are because it's uh processing the resolution on device in addition to the transcript. But once we stop recording, because for for now we this this recording will go through panora processing now. There's there's two ways the recording can go. So the first way is if you're privacy focused and it's a very important thing that your recordings and your information stays private. We have a pipeline that allows for on-device transcription and diarization. And this will run on your phone as long as you download the appropriate models for that. As in the app does it for you automatically, but it will take some time depending on the size of the recording. So it it it does work up to maybe about four speakers. But for now, we use Panora Processing, which is the service that we have that runs in the back end. And how we do that is when the recording is done, we will encrypt it and chunk it up and send these chunks to our back end, and then the backend will have uh we'll decrypt it and process the recording using a powerful uh processor. And once it's done, it will send it back to the phone, but it only sends back the transcripts and the artifacts that we create. These artifacts are like topics, uh people, events or decisions that it has lifted from the transcripts itself, and it also sends back send back the direction uh details. So, like for instance, it knows when speaker A was speaking, and it will split the transcript up into like when Anthony was speaking and when Kitsune was speaking. And this gives the transcript a bit more context as to like who's actually making the decision. Right? Yeah, but because you don't we don't we don't track uh your names or anything, so unless you actually mention it explicitly in the meeting on or in the recording, it will just give it attach a random name to whoever that it thinks uh belongs to a single person, which you can then update or edit yourself, and those those stuff will stay on device. So after that, once that is done, because of the artifacts that we've created, you can use the chat function where you instead of uh reading through the entire transcript yourself, you can just ask directly like uh what did Anthony say in the first five minutes? Uh what was Anthony's decision on something, and it will go through the transcript and the artifacts and lift up the appropriate information and present it to you.
SPEAKER_00Nice. And yeah, I guess so user records a conversation at the end, it gets a summary and it gets uh the transcripts and all the data that it can uh the user can prompt later on. So within Keystone, that's basically how how I see it. I can see the files, and I would also be able to have a chat function to chat with uh all the data that Penora has uh collected, right?
SPEAKER_02Yes, that's correct. The data will live on your phone, and in the settings, there's an option to move it to a folder that you can access, and you can then see what kind of artifacts have been created, and you can see the actual recording clip, and this all stays on your phone. So that I mean you have full control of your data. The our philosophy is that we we do not want to. I mean, first of all, we do not train on your data, we do not take your data and do other stuff with it. All we do is to process it, and even through the processing, it is based on uh time-based encryption key. Everything that is done here is intent-based. There is nothing that uh there is no process that will trigger at the back end that is not without your intention. How do I say this? So for instance, if you are choosing to record and to upload, that is your your intent, and then we will process that. If you are choosing to uh ask a question, that again is your intent, and then only then will we process that. But once the initial processing is done, there is no additional processing.
SPEAKER_00Okay, understood.
Chat With Your Transcript And Artifacts
SPEAKER_00And let's say I have multiple conversations with uh like Kitsune, and for all the conversations we have ever had, I also do record them. Uh are those
Voice Tagging Across Multiple Conversations
SPEAKER_00do you guys have user profiles uh with help within the folders in Panora per user that I speak with?
SPEAKER_02Uh so yes. So one one good point about uh using Panora processing is that when we are doing the dynamic, we also can create the voice embedding of the specific speaker, and we will send that back down to the phone. And the next time you have a conversation with the same person, it will suggest that that person to be, I mean you will suggest the name of that person to be tagged to the to the voice. It's like a voice print, but a more rudimentary, simple way of saying it. It's just it recognizes voices if you opt into it, right? You don't have to opt into this. And that will help you uh compile all the recordings with that specific person in under his uh profile name. So within like if there is a profile for Anthony, you can see all the recordings that Anthony has appeared in.
SPEAKER_00Okay, understood. And yeah, I maybe maybe I can also ask uh for from a user standpoint using on-device versus cloud processing, uh besides the privacy uh being yeah, besides privacy being a feature for on-device, uh are there any uh downsides to on-device
On Device Tradeoffs And Phone Requirements
SPEAKER_00processing versus like cloud processing?
SPEAKER_02Oh, definitely. So in in the office we have a range of different phones and and obviously, I mean, you know, the latest phone is gonna be the one that processes it the fastest because of the newer hardware. On the old phones, they do they do can they still can process it. It it takes probably two or three times longer, and the phone does get warm. But the point I just want to drive across is that if you allow it to work, it will actually go through and work through it as long as it has enough RAM. And so our minimum requirement is at least the iPhone 13. Understand.
SPEAKER_00Yeah, another thing.
Long Term Memory With Semantic Search
SPEAKER_00Um let's say I had conversations stored up to like multiple months. Um maybe yeah, maybe you can also help me understand how it has such long, long-term memory, uh, given that uh like LLMs have uh very short context windows, right? How does it ensure like it remembers down to every minute or moment of conversation that we we've had?
SPEAKER_02So the way we store the information, we will also take note about the recency and the similarity to your question. Right? And whenever there are decisions, they they also come tag with the recency. It's a it's a it's not a like a vector embedding of the conversations that you have because that will be too expensive to run on device. Yeah, so it is a simple semantic search on all there will be an index of all the related documents, and then from there, as the LOM calls the tools, it will slowly read through and discover if it's relevant. If it's not relevant, you will just go to the next one. So it has a certain budget in order to figure out like which documents are more relevant and more recent to your query.
SPEAKER_00Got it. Okay.
Pricing Status And Early Access Slots
SPEAKER_00And yeah, maybe maybe we can touch on subscriptions and revenue streams for Penora. Like what are the subscription models like? And yeah.
SPEAKER_02Right now there is no intention for any subscription or revenue streams. We are still working on that. Right. So so for for now, the first we we have a couple of slots uh, but it's like the first 200. The first 200 of them, you can get full access to everything, to all the features, and yeah, I we will have to update you on that.
SPEAKER_00Okay, understood. No, no worries. And yeah, maybe
How AI Helped Ship In Weeks
SPEAKER_00we can dive into um yeah, the building of Panora. Um how how much of uh developing Panora was done by uh AI? And maybe we can also go down into uh yeah, how do you use AI to build Panora?
SPEAKER_02So so to me, I find AI to be like what's in to quote what everyone is saying is uh productivity multiplier. You you can have an idea and AI can iterate with you, but to actually build an app, there are certain core fundamentals that you need to have in place in order to ensure that it doesn't break the first moment like someone else uses it.
SPEAKER_01Right AI is very supportive. Uh how do I say this?
SPEAKER_02It was instrumental in in in allowing us to ship the features we wanted uh fast, right? The entire development of this took about six to eight weeks from idea conception to to an actual MVP. And through the designs, the the back-end work, a lot of the stuff was uh assisted by AI.
SPEAKER_00Okay, got it. Um yeah, I guess um yeah, maybe you can walk us through like each part of the development, right? How uh what what AI you used, um how how how they access you. Uh are you using open claw as well? Or was it just straight out of uh clawed claw code?
SPEAKER_02Yeah. Oh I I don't have a specific tool that I'm not a fan, I'm not a fanboy of any any one of the big AI companies. I I use whatever is available and whatever I see is capable of doing. But it's not always just a single AI, like a single prompt to a single AI, and then you take and run with the code. No. I normally will have at least two or three stages of reviews from the different AIs and uh look at the code myself before I actually uh commit to it. Right? Because many AIs, especially on the benchmark, they only test for like uh single single shot prompts. But when it comes to a large code base, you need to also allow it to understand the nuances of the actions and the stuff that is gonna affect when you make changes. And if you use the same AI for everything, they are only gonna be very narrow in scope
Model Choices Reviews And Cost
SPEAKER_02and only see certain stuff, right? And sometimes you need to mix it with bigger models so that they have a larger thinking budget to actually consider and anticipate what can or might happen if this, like, for instance, this function changed. Yeah.
SPEAKER_00Got it. Um maybe you can share some of the models that you use uh throughout the process. And yeah, you mentioned having like besides the uh having AI to code it up, you have them to review as well. Maybe for reviewing, are there a certain set of instructions that you provided?
SPEAKER_02Uh not necessarily. The review is more of giving it enough context to understand the problem that you're trying to solve and then letting it figure out whether the solution presented actually addresses the problems. So at least that's my style of doing it, right? Like if I want to create a blue button, let's say in the middle of the screen, right? And so the task is to create a blue button, and I I prompt an uh AI to get the the code for the blue button. I was the blind, the reviewer will be blind in the sense that I will give it a different perspective of the task itself, not the actual like prompt used to code so that it doesn't fall into the bias where it's limited by the initial prompt. This allows it to explore other avenues to see like whether the prom whether the result is actually reflective of the prompt.
SPEAKER_00Okay. And yeah, going back to the like what AI models do you use?
SPEAKER_02Oh, uh multiple, like uh GPT 5.3 to 5.5, and throughout this time we had uh Sonnet, Opus, and now we have Fable, and then when DeepSig came out, we also use DeepSig and GOM. GOM 5.2. Yeah.
SPEAKER_00For GLM you run in or uh no, unfortunately.
SPEAKER_02It is too big to run locally.
SPEAKER_00But DeepSig, DeepSig. Deepsig is okay for locally, right?
SPEAKER_02DeepSig is about 1.6 trillion parameters, so that's requires about 600 to 1 terabyte of RAM. Oh yeah.
SPEAKER_00Okay. Um do you see like which models uh were better at actual coding uh or features? Or yeah, maybe maybe you can let me know uh from your experience which models are advantageous in in what uh specific uh use cases.
SPEAKER_02I believe the model and is important, but the harness is also important in trying to get like uh trying to Understand the prompt. So like you can use uh GPT in Codex, or you can use GPT in open code or Pi. And the completeness of the work is slightly different. And the time it takes to complete the work is also slightly different because Codecs is quite bloated. Pi is script down, but it also means that its scope is very narrow. The one that really impressed me was uh DeepSeek. Especially for its size. It can come up with a solution that is almost like 90% compatible with what Opus and GPT 5.5 can give me.
SPEAKER_00Wow, yeah, that that is shocking. Especially with the cost differential, right?
SPEAKER_01Yeah.
SPEAKER_00Um yeah, and yeah, actually talking about cost, uh do you guys have an estimate of how much uh tokens were spent? Or like how much do you guys spend on tokens?
SPEAKER_02Probably more than 200 million.
SPEAKER_00Okay. But I mean, uh in in in the sense of developing the app, right? Uh would you say that uh AI was essentially practic like maybe an uh a software engineer intern uh equivalent?
SPEAKER_02Yeah I would say it's halfway there. You still need some you still need someone to guide it, especially in terms of like taste and the direction that you want to go. It is more of a tool. If you use the tool properly, you will get the results that you want. If you if you give it too vague a rat uh prompt, you you probably end up with slot. Yeah.
SPEAKER_00Okay. And regarding the like you earlier you mentioned agent harness. Um is that is that something you customize yourself or uh are you just using off-the-shelf harnesses?
SPEAKER_02Uh we try to use mostly off-the-shelf harnesses because we work on different systems at any one time, so it's a bit of a logistical nightmare to try to ensure that each system has got the exact same setup. Right. I trust the model to be smart enough to determine. But the harness will help me with the creature comforts. Like making sure like the linting and the what else? The completion of the problem. So like when you give it a task, right, sometimes it doesn't complete fully because it achieves the first the first task and it thinks it's that it's complete. But the harness will actually be there to remind or to steer to ensure that the AI knows that there is additional tasks to be completed in order to actually say that it's complete. Okay.
SPEAKER_00Okay, got it. Then yeah, um let's say let's say if uh other people are looking to build a mobile app and they're gonna use predominantly AI. Uh would you say I guess it's it's possible? Or yeah, or it you do you still need essentially software engineers to guide it through the
Distribution Challenges And Go To Market
SPEAKER_00process?
SPEAKER_02Uh okay, I really think it depends on the type of app and the scope of completion that you want the app to be. If it's a simple app, like just like if you want to make your own calculator, it should be quite easy to prompt with AI. But the moment you try to have additional users to your calculator, and you need a back-end service, and you need to worry about load balancing and you know like the processing power, the processing needs of the app, or whether it should be on device or off-device, and how do you control that? How do you control abuse, how do you control uh an influx of people, account management and stuff. All this, you at least need some level of fundamentals on what it takes to uh set up the backbones of this before you actually embark on it. Because if you just completely trust AI, they are gonna use the packages and the tools that they have been trained on. And sometimes it might work for your situation, sometimes it might not. And well, you can't get too attached to it anyway.
SPEAKER_00Got it. Um nice. Yeah, thanks, thanks for sharing about uh how you guys used AI. Maybe we can go back into Keystone uh and also yeah, a bit of Panora again. Um regarding, I guess, yeah, building a Panora, were there any challenges that you guys faced while building this?
SPEAKER_02I mean with the current landscape, the challenges are mostly uh distribution and trying to get your app like legitimized on the app store. There's certain hurdles that you need to jump in order to get your app published uh easily accessible for users. Apart from that, the building part is it's much more convenient nowadays. Yeah. So from ideal generation to a product, it's doesn't it's not gonna take as long as it used to before, but you're gonna be stuck or at least burdened by the reach of the distribution. Like it can be useful to you, but whether other people deem it useful to them and to actually use it in their lives is a totally different question. It's like separate from building.
SPEAKER_00Got it. Yeah, so I guess at the end of the day, uh I mean, good thing about the current age of uh AI and coding is that yeah, building it's easy. Uh but the bigger hurdle is still distribution and actually uh getting getting users. Good point. Yeah. Um yeah, maybe I can yeah, so get a better understanding for for you guys, uh given that you guys just launched. Um what's the plans around go to market and uh marketing uh Penora to get it in the hands of the first hundred users?
SPEAKER_02So right now we are actually at about 15 users, and I think that's pretty good for day one. Um we have ads running and we have uh marketing and we have all those posts on social platforms, but the conversion rate is not as high as we expect it to be. So I think we have to look into that to see how we can uh improve that.
SPEAKER_00Okay, and yeah, you also mentioned earlier the difficulties of uh getting an app on the app store and also app app rankings, right? To to get it such that uh users can even see see your app. Um yeah, are there strategies that you guys are looking to implement to get the app on a higher ranking? Yeah, yeah, making it more visible in the app store.
SPEAKER_02I'm not so sure though.
SPEAKER_00No, all's good. All's good. Yeah, I just started. Um and uh yeah, for the next part for Panora's roadmap, maybe. Uh what given that you guys just launched uh let's say further down into Q3 and Q4, are there any potential new k new features or new integrations you guys are currently uh looking at?
Roadmap New Products And Scope Creep
SPEAKER_02So right now in Keystone Lab, uh how do I say this? We currently have a second app in mind that we are a second app that a product that we are developing while we uh slowly allow Panora to get a bit more FaceTime and see whether the traction increases. So the next app that we are building is something that uh we hope pet owners will appreciate. Or even uh maybe not pet owners, but anyone who pets for animals, yeah.
SPEAKER_00Okay, yeah, that's interesting. I think uh I think animals are quite viral on TikTok in general. Like even if you don't have a pet, people still love cats and dogs. Yeah that that should hit. It's yeah, yeah, maybe you can go back into Keystone as well. Um what's the process for you guys uh that informs you guys on what apps to build? Uh how do you guys get inspired? Uh are you guys looking at trending uh items uh or trending use cases or like increasing total addressable markets? Uh yeah, maybe I can get just just an understanding of the principles behind how you guys uh decide what app to build.
SPEAKER_02But now we don't really look at trends, we look at problems that we want to solve that no one else is solving, right? Uh maybe gaps in the current solutions that are out there, or if something that uh is monetized, but actually should be monetized differently, right?
SPEAKER_00Okay, yeah, that's a that's a good angle. Yeah. Um yeah, apps that people yeah, that solve gaps in the market. Yeah, that makes sense. Uh and yeah, it's uh maybe yeah, we're down to my last few questions. Uh given that you guys just launched this app you're and you're already developing the next one, is there a goal in the number of apps that you guys are looking to launch for the rest of the year? Yeah, is this more of a is this strategy more of a numbers game in this case?
SPEAKER_02Uh right now we are looking to launch like one product every quarter.
SPEAKER_00Okay. Got it. So I guess we can expect one more in Q4.
SPEAKER_01Perhaps.
SPEAKER_00Nice. Okay, okay. Yeah, um yeah, maybe just to round that up. Uh on your side, what what are some key learnings you had from building Panora? And uh yeah, how are you translating that to the second app? And yeah.
SPEAKER_02I mean, the first thing you learn is that scope creep is real. Scope creep? Yeah. So when you create you when you have an idea for an app, and then you rationalize and imagine the level of uh functions and improvements that you can tag onto it, and then you realize that the more you think about it, the more it can be done differently, and then your scope of work becomes like larger and larger, and then you find yourself in a place where suddenly you'll you want to have too many features on the app where and time is like constrained.
SPEAKER_00Yeah. I think that's that's something most people who use AI face as well, right? Uh when you're coding with AI, they always suggest new features. And they usually give like three or four recommendations of what to add.
SPEAKER_02Um not just like AI. Sometimes through actually like using the prototypes, you realize that the it can be done slightly differently, but the the ideal way that you wish to do might actually take up much more work than you anticipated. Yeah.
SPEAKER_00Got it. Yeah. Actually, this yeah, this is more personal. Given that there's so many potential pathways to take, right? Like different features, different ways to do the current features, even. Uh yeah, sometimes I wish there is a way where instead of just making a decision and hoping it works, I can like A-B test, you know, either either variant.
SPEAKER_02Yeah, then you effectively doubled your work.
SPEAKER_00Yeah. I think that's just infinitely more work. It's like a Hydra. Never end. Yeah. But yeah, it's just interesting to think, because uh building apps, yeah, yeah, you're right. Time, time and effort uh plays a huge part. But sometimes I do wonder like some of the paths I didn't take, uh where where those could have led, right?
SPEAKER_01Mm-hmm.
SPEAKER_02So you gotta make a decision and I mean accept the compromise. Or at least find a way to live with the compromise.
SPEAKER_00Yeah, that's a good yeah, that's a good point. And yeah, uh, we are coming to the end of the podcast.
Where To Try Panora And Follow
SPEAKER_00Uh maybe yeah, you can have some final words for Panora and Keystone. And uh yeah, let us know where your socials are so we can uh also look to uh give you guys a follow and yeah, follow your channel as well.
SPEAKER_02If you want to try this meeting app, you can always go to uh penora.keystone.ai. We currently have like 185 slots more for the full premium access.
SPEAKER_00Okay.
SPEAKER_02Yeah. Social Spice. Uh Social Spice is uh Panoraai at underscore app for Twitter and Instagram. And um Keystone. Yeah, I think that's about it.
SPEAKER_00Keystone Key. Okay, yeah, that's great. And yeah, thank you, Kitsune, for your time today. And yeah, uh that wraps up uh today's episode.
SPEAKER_01Alright.
