# Neat's AI Strategy for Proactive Meeting Rooms

**Podcast:** Chain of Thought | AI Agents, Infrastructure & Engineering
**Published:** 2026-08-27

## Transcript

The meeting rooms where people get together and typically in the cases where there's also somebody remote and how can we make that the best possible experience?
Eleven or more cameras in some cases, we have dozens of microphones.
Who's present?
What are they trying to do?
Who's looking at who?
Who's talking?
The meeting room transforms into sort of proactive participant in a meeting that it bends into the background.
You make the meeting and we make it work.
My guest today is bringing AI into the meeting room.
So let's talk about what that means.
Welcome back to Chain of Thought, everyone.
I am your host, Connor Bronsden.
If you're getting something out of these conversations, or if it's your first time here, take a second to like this episode, subscribe, comment, and if you're on YouTube, turn on notifications so that you actually get the next one when it lands.
But I want to say a quick thank you to our presenting sponsors for season four of Chain of Thought.
Sfix, delivering reliable enterprise webhooks at link.sbix.com slash C-O-T, and Walrus Memory, giving AI agents portable, verifiable memory at walrus.xyz slash C-O-T.
That guest that we are chatting with today is Tormid Rhee.
Tormid is the chief product and engineering officer at Neat, the Zoom-backed video hardware company.
Before Neat, he co-founded Ava, a computer vision security company.
that Motorola acquired back in, I believe, 2023, and spent close to eight years at Cisco, where he ran the Spark board.
And I am certain we're going to be talking quite a bit about computer vision and what it means for computer vision models in the years to come.
So really what we're getting to here is that Tormund has built both halves of this, the meeting room hardware side and the computer vision that reads the physical space.
But how does that all come together?
Tormund is here to tell us.
Tormund, great to see you.
Thank you.
Thanks for having me.
It's great to be on.
Yeah, I'm really excited to have you on Chain of Thought.
We don't have enough conversations, I think, about how AI is impacting physical spaces.
And one of the key areas of that's happening is that as we are in this geo-distributed world with teams all over the world with customers, partners, the meeting room experience has really changed the last few years.
And I know Neat has been at the forefront of that.
What does it mean to turn a meeting room into actually a system that understands the kind of meeting happening and actually directs the experience rather than just framing whoever is talking?
That's a good question.
I've been in this industry for quite a while, since like around 2010, 2011.
We tried a lot of things to get people using that technology, including Icelandic volcanoes and arguments about travel and cost reduction.
I think the pandemic and the boost we saw in usage after that has really changed things, definitely.
However, to some extent, the technology that people are using now for the distributor or the kind of hybrid workforce or call it whatever you want is typically built pre-pandemic and pre-ubiquitous use that we see from it now.
So to date, I don't think you've properly used AI or like one thing, it's sort of to some extent built with sort of in the past.
And to another extent, there's so much happening in the AI space now that hasn't really made it into the meeting room.
So that's really our job.
It's to take the meeting rooms where people get together.
And typically in the cases where there's also somebody remote and how can we make that the best possible experience?
And in order to do that, we actually need to understand what that room is like.
and what the people in the room are trying to do.
Because ultimately, we want the technology to just move into the background and for people to focus on the conversation and the meeting and just have something that just kind of works and makes the interactions simple and natural.
So what does it take to actually understand what's happening in the meeting and then drive these outcomes?
Yeah, quite a lot.
But ideally, you don't notice.
So like the objective is for it to be very simple or transparent, but there's quite a lot of like complex or sophisticated technology behind it.
So for us, it means having a pretty kind of broad set of sensors in the room and like the most important sensors are cameras or microphones.
And then we do have some others for things like air quality and other things, but it's typically microphones and cameras.
We have then like 11 or more cameras in some cases.
We have dozens of microphones.
We use them to build sort of a spatial model of the space, but we also use them to sort of understand what's going on or what we like to call reader room, but it's basically figuring out like who's present, what are they trying to do, who's looking at who, who's talking, is somebody standing up, is somebody presenting, what's the interaction like, looking at sort of both verbal cues.
non-verbal cues.
And there's been a lot of progress over the last few years in sort of the detectors, like you can detect people quite easily.
You can understand who's talking.
But sort of the next level signals, like non-verbal cues, for instance, are combining these signals from multiple cameras and sensors and making that work like a system, not just from sort of one powerful device at the front of the room, but across a set of devices in the room.
That's the next hard problem.
And the one that we are, yeah.
focused on solving and making improvements in.
So quite a few of these models that are helping to read the room are running on edge.
Is that correct?
Yes.
So we do a combination of edge and cloud.
We sort of do it where it makes sense.
But typically, sort of arguments for doing it at the edge, probably in order of importance, is kind of time sensitivity.
privacy and cost and cost tends not to be the kind of main factor but but like anything we do in the media path anything that has to do with processing audio and video which is a lot of what you have to do when you're kind of trying to kind of foster a meeting because it's about like seeing and hearing people needs to happen at the edge so that all happens at the edge uh for for good and bad and things that are less time critical things that have to do with like management or maybe things that you can sort of do post-meeting, like post-meeting summaries, transcripts.
Those things can happen at the edge, but what needs to be done to produce a great meeting needs to happen at the edge.
So are you largely using like a combination of open models that you fine-tuned for specific tasks or are you building models in-house?
What's the approach that needs taking?
Yeah, it's a mix basically.
So we sort of...
Some is open source, some we build in-house.
We don't reinvent the wheel.
I think the wheel differentiation isn't so much in the model for us.
It's sort of in the harness around it, which means like how do you combine signals from different detectors or different models?
How do you make the right decisions?
How do you make sure that that works in a real environment?
So there's a lot of variability in sort of room acoustics, sizes, how people behave and interact.
And in sort of that case, we work across the stack.
We work from data collection, labeling, training, on-device inference, obviously, and then the harness around it.
So, yeah.
And then I think also like model capabilities potentially commoditize over time.
So it's like the harness and the specialized data is where we have the magic sauce, so to speak.
I would love to understand more about your approach in the data piece before we talk, I think, more explicitly about the harness.
Can you tell me a bit about how you have architected your system so that you have this data advantage?
I mean, obviously, you talked about the sensor piece of, look, we have 11, sometimes more sensors in a room at times.
So that's an advantage.
But how does that all come together in a way that you can effectively feed information across these models?
A good question.
I can do some of the high-level stuff, but I can't get too specific because this is a race where you're trying to keep something a bit of a secret to, obviously.
Data isn't like, data is not an easy, I guess it tends not to be an easy problem, but our devices are in meeting rooms where people have important and private conversations.
So we can't just collect data from all our meetings and our customers typically wouldn't be okay with opting in to data collection either.
We sell to It's a B2B company, so we sell to smaller and larger enterprises.
And it's very, very important that we respect the privacy of what goes on there and what happens in the meeting room.
So what happens locally stays locally and is processed locally, which means kind of getting good data is a bit more of a challenge, but it's a combination of internally sort of source data, publicly available data sets that are more...
relevant, not that much synthetic data, kind of have looked at that a bit and yeah, sort of a data set we've built over time to make that effective.
I love that you brought up the privacy element because I think that is obviously crucial for so many businesses and maybe it's ignored at the demo stage, let's call it, of folks trying this stuff out.
How do you set...
the guardrails that you need between all this data collection and then the training you're doing with it before it gets into how the like direction of the room and the harness beside works, the coordinate.
Yeah, we basically don't collect any data from our customers that have to do with sort of no audio, no video.
So that needs to come from somewhere else.
So that's a very hard line that we draw.
We also...
process things in like so so we don't use their data to train our models we need to get it from elsewhere um we also uh tend not to deal that much with identity which is like another like we we don't we don't need to know who it is right so we don't need to when you like for some things in the future that could be useful but like typically we don't need to know that it's Connor or that it's Tormod we just we just need to know that sort of that person is the same as that person and we sort of look at the interactions and how they work together in the room but identity at least has been less important for the experiences that we produce.
We try to just, I guess, infer importance or relevance from all the other things than identity.
This is interesting because it, I think, speaks to the importance of the edge model approach that you're taking too, where you're able to localize a lot of this decision-making, be able to keep that data localized, instead of saying, oh, we're pushing this all to a cloud, which then creates all these governance issues, these privacy issues.
I love that you have this structure here, but it does sound like you're using the cloud for some things.
Is that kind of the director or harness piece that is using cloud models?
And then...
Basically, there's a line between how that works or how does the actual coordination piece work?
Because I know you have all these different pieces of the room that are working in concert together to deliver this great product.
Yeah, good question.
I think then it's probably important to mention that so the devices that we create, they look a bit like these bars or these boards that you put into the meeting room and you use them to have Google Meet, Microsoft Teams or Zoom meetings.
And we also support some third-party applications.
that has to do with the sort of meeting, that sort of media, audio, video is done at the edge, which means that the kind of orchestration also happens at the edge.
When we have multiple devices in the room, we might have, and they all, we sort of distribute to compute.
So in the room, there could be multiple devices.
There's like a main device that's at the front, typically.
has microphones, cameras, speakers, like it powers the screen, things like that.
Then there could be like what we call a need center at the table, microphones and cameras.
And you sort of have these devices across the room.
Then they all do their own edge processing on audio or video or both.
And then they send a combination of like metadata and media to...
to sort of the main device.
So it needs to sort of be this kind of captain main device in the room that does the level of orchestration.
But that job is easier and lighter if the other devices in the room sort of do their part and send metadata along with the media, basically.
So then it can make decisions from looking at the metadata and it does not have to do the inference on the media streams that are coming to it.
And it can do sort of more like selections or mixing or you can call it pick.
this video stream from there and that one from there and kind of combine them together or it can kind of mix the audio appropriately.
But both then sort of the inference and sort of the harness and the orchestration in our case happens at the edge.
However, these devices are used to have meetings on, as I said, like Google Meet, Teams and Zoom.
That's obviously a sort of a cloud meeting service.
So when you look at the end-to-end solution, then on their part, there is definitely an element of cloud processing.
So that's the meeting part of it.
And then you asked about, like, what do we then do in the cloud?
It's lighter than what we do at the edge.
So it's typically, when it comes to AI, it's sort of...
I guess we're a bit more of a, we have a cloud management suite as well.
In that case, we're a bit more of like a management platform.
So in that case, we could use AI to make the management platform more effective.
But it's also, I think, increasingly sort of a platform for integrations with like events, webhooks, APIs, and an MCP server for agentic type management.
And in that case also, we're not really doing, we're not sort of the agent, but we're sort of the platform that you interact with.
And then there are some examples where we, in one of the modes that our devices run in where we do AI type summaries for the audio.
And yeah, currently for the audio, I'll leave it at that.
And that happens in the cloud because it's slightly less time sensitive.
We process the audio either during or as the meeting closes.
That's a long answer, but that's what we touch and what we do there.
I love the detail.
And I think it's...
really interesting to look at this from the standpoint of like you are coordinating a variety of edge models across hardware with this captain and i think increasingly we see this model in every style of ai where whether it is uh you know someone coordinating a team of agents to do coding whether it's you know a meeting room in this case of having this orchestration model this leader that is running a group of sub-agents and i i think that parallel is one that we're going to see across the board.
It's not always going to be the fable of the group that is doing every piece of task.
It's going to assign sub-agents to go out and do stuff.
And it's interesting to hear that that's happening on the hardware front as well.
And I'd love to understand some of what this actually brings back to an organization.
So obviously the meeting summary piece you talked about, there sounds like there's some great options there.
But what does the agentic platform that you mentioned and the work you're doing to provide an MCP, et cetera, so that people can leverage their agents to interact with the meeting documentation or a meeting organization, what does that look like when it's successful for an organization?
What do they actually get out of that?
Yeah, two couple pretty distinct type of users in our case, right?
One type is the people that have the meeting, that go into the meeting room or whatever space that device is in.
And for them, it's supposed to just work and be incredibly simple.
People are fed up with all the complexity around having meetings.
And there has been a lot of progress, but the majority of the progress has been typically in the small to medium-sized rooms.
You can put in a bar, it works, you have a good experience.
But it can still get complex in some cases.
It tends to get complex if you have larger or more complex spaces.
think like large meeting rooms, auditoriums, training rooms, those kind of spaces, where you still quite often have to fiddle with like camera presets or somebody has programmed microphones to trigger the panted zoom cameras.
And then you lean back and you're outside the camera view.
And it's sort of, it's like, it's like, ah, that doesn't feel right.
You can do this better if you just, if you actually understand what's going on there.
Yeah, and then there's like the interactions that people have with the systems in the rooms.
Like, should you really expect people who they are to have a discussion to have to deal with like camera views?
Like, should they have to click a button and have like a dropdown of five different selections for the kind of framing that they want?
They shouldn't have to do that.
Like they should just, it should just adapt and understand the conversation and the meeting and then just like let the people in the room just go on with like, having the meeting and discussion, focus on discussion, let the remote participants feel like properly included in the conversation.
So that's like one problem.
It's the users.
It's become a lot simpler, but there's a ton more you can do.
And AI is critical.
As you said, like it is about orchestrating AI across all these devices, which makes it an interesting technology problem.
It has its benefits.
It scales more kind of flexibly and et cetera.
But like that's one problem.
And then the other problem you're looking to is like, that's the IT that's admin problem, right?
It's like, I now have like, sometimes a few, sometimes tens, hundreds, sometimes thousands or tens of thousands of devices.
How do I manage all these?
I have various different like management tools that I can go into.
This single pane of glass thing that I've been wanting for the last 20 years never materialized.
And I keep having issues.
How do I stay on top of things?
On that, That IT admin side, I think where it's like agentic management will be very important where people will stop going into kind of logging into UIs and clicking around and kind of looking for red lights to fix things.
And they will have agents acting on their behalf in towards our platform and in towards other platforms too.
Like for the meeting platform provider, Swim Teams, Google, Network, Building Management, who knows what.
That enables them.
just to do what they're doing quicker.
So they can just kind of have like an interface where they can describe or say what they want to do.
They don't need to know the commands.
They just like, they describe intent basically.
And then the agent acts on their behalf.
But I think even more interestingly, the agent can then monitor, manage, and to some extent, like proactively resolve issues.
So this one has a problem.
Should I look at the network configuration?
Should I reboot the device?
Should I do X, Y, Z?
Somebody opened a ticket for an issue in room X.
Let me go check on the status.
Somebody sent a Slack message about poor audio quality in room Y.
Let me go have a look at what's going on.
That's definitely going to happen too.
And the first one is sort of on us.
The second one is, I think, will be about providing an open platform that allows our customers to do this effectively and sort of having the relevant, what should I say?
AI interface.
for it.
And that's like, we've seen just so much progress there just over the last few months, to be honest.
If I talk to some large, I remember I had a conversation with some of our large customers late last year, asked about AI, using it for other things.
Topic was still, yeah, that's interesting for providing, for producing a better meeting experience.
And now they're kind of, they're all part of the MCP beta and everybody asks about integration with different.
different agents and that's like this made made a ton of progress just in a few months so that's very exciting to season four chain of thought is delivered by sfix we spend a lot of time talking about what agents need in production and one of the least glamorous answers is events your customers want agent workflows that react to things happening inside your system which means your api needs webhooks that actually work not just a post request and a prayer retries ordering idempotency Replay protection.
Sfix does that as a service, and they wrote standard webhooks.
The spec that Anthropic, OpenAI, and Google bailed against.
So if your API doesn't have reliable webhooks, that's turning into a lost deal.
Join Brex, Drada, Daytona, and many others on Sfix.
Get started at link.svix.com slash C-O-T, or go to the show notes to grab the link.
Qualified startups will get $12,000 in credits, $50,000 for YC companies, I can't recommend Sphix enough.
I'm a huge fan of their open source project.
I've actually contributed a bit myself and they're so easy to integrate with.
I think you'll really enjoy it.
Check out Sphix at link.svix.com slash C-O-T.
Yeah, it seems like that is taking off and there's so much more potential there because it's pretty clear to me that Neat has made this bet that the next generation of shared space hardware is not just the hardware but this application platform that you're building on top of it to enable organizations.
So from meeting device through that platform there, it's clear that we have more potential to take advantage of it.
Even if we don't improve the hardware in the short term, which I think Neat probably will be doing as well, we have a lot more that we've built on the top of it.
But there's not like an app store for the meeting room at this point, I would presume.
There's a neat app store, actually.
Oh, never mind.
Okay.
This is an interesting challenge as well.
So customers buy our devices and their enterprises, and they typically expect them to last for at least plus minus five years.
Quite often, typically five years.
So they buy them.
It's a long-term decision.
they wanted to be relevant and useful for at least the next five years.
And then we have all this change happening in parallel, right?
And now they're on PlatformX.
They might move to something in the future.
There will definitely be a new application that you want to use in the shared spaces where our device is in that we haven't even heard of yet.
When you have these major tech transitions, there's always like a new player or a new application.
that's like relevant for how people work together.
It's like too early to tell which one it is now.
We don't really know, but I'm pretty sure there is one.
And then some of the incumbents, they kind of move quickly and they fast follow and they catch up.
And some of them like sometimes fall by the wayside.
And they're like, if you look at previous transitions, you will prior to the transition, see names that are no longer here.
So then we have to, we still have to like provide something that gives our customers comfort that that device is useful.
for the next six years.
So the way that we're doing that is that you can use our device across all the main platforms.
We also have what we call neat app hub that allows you to install and run other applications.
Like it could be for like collaborative whiteboarding, project management, also use cases outside of the meeting room, wayfinding, visitor management, things like that.
So it's useful for more things.
It's like it's a platform to enable other applications.
And then we make sure that the ecosystem is open.
I think quite a lot of vendors, when they see AI, it's like, okay, how can I make a new AI feature?
And they add a button or a capability or something within their platform.
But AI and agentic workflows now is just as much about working across platforms.
So you need to make sure that you're...
platform is open and it's easy to integrate with and then sort of it's always been the case for NIT so that used to be like a sort of a broad modern API surface events, webhooks now it's an MCP server so it's just making sure that we continue to evolve that in a way that's also useful for agents so that's like our keyword for that is open any platform, any applications any integrations We're here to be open to support and run the applications you need to run and to integrate with whatever you want to integrate with.
And quite often now that's some agentic workflow.
But yeah, that's important to us.
How do you expect these agentic workflows to evolve over the next year or two on top of the Neat platform and through this open approach that you're taking?
Yeah, that's a good question.
I think a lot will happen in the meeting room, but we sort of own that.
So we have to do that.
Like we have to then have new detectors or use agents to kind of produce better meetings and sort of be the director for the meeting room.
When it comes to the product management side of things, I think we're going to see kind of a pretty rapid adoption in...
using agents to effectively manage device fleets also at scale.
So anything that has to do with monitoring, anything that has to do with management, anything that has to do with installations or configuration.
And I think, again, it's going to move into...
It's not self-healing.
It's like, does anybody, that's probably a buzzword, I'm sure, agentic healing.
But like, it's going to be an agent that sort of then does actions to resolve issues, both directly in our platform, but also across platforms.
That's the next year.
And then who knows what happens after that.
But I think that's definitely going to happen.
That's happening now.
So that will just accelerate.
And then as far as in-room.
Obviously, you mentioned detectors as an area where there's improvement opportunities.
Models are a clear opportunity as somewhere that you've been leaning in and needs training their own custom models, as you mentioned earlier.
Can you talk a bit about that interim experience and how you're thinking about improving the models that you're using on the edge here?
I know you can't share too much detail, but we'd love to understand a bit about the decisions you're making around.
When to use custom built and trained in-house computer version models versus pulling from exterior sources and why you're taking the training approach that you are.
I think we just flexibly use whatever is best.
If we can find something that's available that we can use, we'll use that.
If we need to build something or customize something more, we'll do that.
I don't think we have like a, in this case, we'll do that.
In that case, we'll do that.
Typically models are trained on data.
It's a bit odd, right?
Stuff that goes on in the meeting room.
It's not like a security camera or a car or stuff you typically find online.
So there is an element of making sure that that model is trained on a relevant data set and tested and works in a real world environment.
So that's just a continuous thing that we have to keep up with.
an area of focus for us now is making this work, like right now, is making that work well across devices.
So increasingly, we're doing like multi-device rooms, especially when they get a bit larger.
Neat initially was like about like a simple all-in-one device.
You bought us what we call the Neat Bar.
You put that on top of your TV.
You bought a Neat Board that also had the screen.
You put that in the meeting room.
Now you're putting more devices into the room.
How do we think about like models, inference, cross-device collaboration and orchestration.
So that's an area of focus.
And then I think where this is going next is like, so today, typically like detectors are like ML or computer vision.
But the decisions that you make to produce the end, the result experience, those algorithms are still to a large extent deterministic, right?
So if somebody does this, or if somebody talks, or if you have more than X people, do this.
That's also going to change at some point.
So sort of the actual director or what makes the decisions about what layout or what experience you want to produce and provide to the far end is also going to be sort of a model that's trained and it's like inference and that's how you make the decisions, that transition hasn't really happened yet.
It requires some amount of data, probably more data than most companies have today on this.
And then it also requires a decent amount of compute.
And because we have devices and we need to work on devices that have been in the market for a while, we pick chipsets that are supported for a long time.
There's also some compute constraints on what's currently deployed.
So that's something that's going to happen next.
And we're going to try to be there first.
The tech and what's in the room is not quite ready for it yet, but that's definitely something I think is going to happen in the not-too-distant future.
It's super interesting talking to you about this because the constraints that you and the team at need are operating under are so different from some of the constraints.
for when I talk to someone who's focused on like AI coding, for example.
For them, it's, you know, hey, I'm just trying to get access to, you know, new GPUs from NVIDIA, from AMD, from whoever, and the hardware is continually evolving.
But to your point earlier about companies expecting that, hey, my meeting room is going to hold up for five years, maybe eight years at the outside, depending on where things are.
There is a much longer lag of...
um hardware capability and so you're forced to operate under these edge constraints and it i think it's making for some real innovation in that space we had a great conversation with maxime laban uh liquid ai last year about their approach to post training and how they've had to think through the differences as far as approaching like smaller post-train models and the gains they can find there.
And I'm sure you're seeing the same because as you pointed out, you have this other constraint too, where you can't train off your customer data.
Like this is enterprise data.
They want to hold that tightly.
And I can imagine that that is creating constraints around what you're able to do in training because there's only so much we need to get out there, as you pointed out earlier.
You can't really train that well, I would assume, off of like a security camera feed.
It's very different from what you're doing.
How are you solving these constraints?
And like, what do you do to successfully operate under them?
I'd love to understand more about like how the team solves these problems.
Yeah, that's like, I think like often like, what's it, resource, constrained resources, like that sort of has to foster, if you feel like you have a box to innovate within, your resources are constrained, what you do, you have to be innovative, I guess.
Yeah.
But you're definitely right.
So our devices are typically Android devices, which you can think of the computer a bit like a mobile phone, typically, right?
So that means that what we have to deal with is the latest generation of mobile compute that was available at the time we designed that piece of hardware.
Which means that if you expect a five-year lifetime, then you'd have to operate with technology that's a bit like what you would find on a new phone five years ago, basically.
At least if you look at a single device, and then there are multiple, but that's what...
So it's not really a laptop even.
It's not a Mac Mini with a lot of memory.
And it's definitely not cloud-level GPUs.
But...
So what does that mean?
Well, it means a few things.
So we do hardware and software.
From a hardware perspective, it means that we need to be quick to market with the latest generation chipsets, i.e.
be close to the chipset vendors and manufacturers, get early access, build devices quickly so we don't lose time and spend two years making the product after the chipset was available.
And then we need to build in enough headroom.
for the rest of what that device has and does, if it's like cooling or memory or the other decisions that we make, so that that device has headroom for future innovation.
So we can kind of run the chipsets to its fullest.
We have some more room to potentially do local models, things like that.
So from a hardware perspective, do devices quickly on the latest generation hardware, build in headroom.
From a software perspective, it's a lot about optimization and efficiency.
which also fortunately gets better over time, right?
How can we do the detectors more efficiently?
How do we do those across different devices?
How often do we have to do it for different things?
And that's like a continuous optimization journey.
And then we have to make that work alongside running the meeting application, which also consumes resources, which also wants to use more AI.
So yeah, I guess it's like a...
Continuous optimization, making sure you have great hardware out there with some headroom, and then it's a continuous optimization game, basically.
And the hardware that's out there now, yeah, well, the hardware that's currently available that you can build products with now is a lot more capable than what was out there when a lot of devices were built in the past.
So that's also exciting.
We'll get to do newer, cooler things with new hardware.
But yeah.
Season four of Chain of Thought is also presented by Walrus Memory.
I spent much of this year talking to guests about agent memory, and we're all having challenges that go beyond simple session handoff.
You explain your code base to one AI coding tool.
You get it tuned to what you need.
But when you look to handoff to a new model family, you find yourself having to confront the same memory and context problems.
Walrus Memory is solving agent amnesia with a portable memory layer for agents.
Walrus Memory takes your agent's memory anywhere you need it.
The context your agent builds can be read back later from a different app in a different runtime, weeks out, and you set the rules for who reads and writes it.
Python and TypeScript SDKs with native MCP support.
I'm glad to have them behind Season 4 of Chain of Thought because this is a key problem confronting AI builders in 2026.
Learn more at walrus.xyz.cot.
Yeah, it seems like there is a lot of potential for this area to grow and for us to have these more integrated meeting experiences.
I mean, obviously, there's companies like Meta that have made big bets on the metaverse and these other ideas, but those seem to maybe be a little farther afield.
And in the short term, there are opportunities around, you know, what does the meeting room deliver?
What can we actually drive here?
And, you know, while we have agents that we're starting to deploy to support this and pull information and align to the data in the room.
I can imagine that there's just a ton more opportunity here and you've talked a bit about it.
Are there particular areas that Neat is planning to invest in around opening up the platform and enabling partners to build applications or to, I guess, what's the pitch for a developer to build for the meeting room?
Where can they make an impact?
Anybody who builds for teams.
So a meeting room is a bit different from an individual use case because it's not like quite a lot of applications are built for an individual user.
What's useful in a meeting room is typically applications that are built for a shared use case.
So it's something that you want people to come together around.
So if you're working with some kind of solution that supports...
or helps a group of people, that group is not always remote in front of their individual laptops.
They're like sometimes together in a meeting room.
If that's what you're doing, then Neat has a platform to deliver that application into the shared space.
Like project management, whiteboarding is like typical examples today, but like with agentic interfaces too, that's going to be different than something else, right?
How, like if you're thinking about, agents supporting groups working together.
I don't know what that could be.
I imagine we'll see.
There's going to be a lot of innovation there.
If you want to deliver that to a group of people that are together physically, then Neat has a platform for that.
It's quite easy to experiment with.
We have a developer platform.
We have well-defined APIs, SDKs, so you can get or pull information from our devices.
We want the microphones or the cameras if you want to interact with the screen.
whatever you want to do.
We have a platform for that.
I think we're going to see a lot of innovation there.
We typically have either standard Android apps that participate in our program and are part of our app catalog.
We support web apps, which is another easy way to just deploy something to our device.
Then we can basically do anything.
And I think also we're going to see companies now having their own sort of company AI built apps.
So there's going to be like an influx of that as well.
So we also want to be a platform for that.
And we're thinking about our role in that.
How can we make it easier to help you maybe create or at least deploy an app that's relevant for your team?
Love that.
I think it speaks to this vision for what's possible and how you're going to be able to enable your customers and partners here.
And super excited to see where need is in a year or two as you continue to have these new capabilities on software and hardware.
Let's talk a bit more about the hardware side.
I can imagine...
getting enough compute to the edge, whether that's NPUs, GPUs, memory, et cetera, plus the right cameras and mic, is a bottleneck.
Where's the gap today between what you want to run locally and what the silicon actually allows?
Yeah, what we want.
Like, we'll just make the most out of what we have, I guess.
A good question.
I think, so, So typical devices, they run sort of these system on chip.
So it's like a combination of CPU, NPU, GPU.
That's sort of a package that you then deploy on your device.
You don't make, you tend not to make separate choices about CPU, GPU, NPU.
You sort of have this version that you can get and it has this combination.
And then obviously, if you look at acceleration on first GPUs, now also NPUs, that's like, if not an order of magnitude, then at least significantly, relatively speaking, significantly more efficient and capable and much more sold in the progression that we've seen in CPU because that's what we need more of.
So for us, that means there's a few things we can do.
We can put more devices into a room so they all don't necessarily have the latest gen chipsets, but we can put a few more in so we have more compute in the room.
So we distribute, like unlike most others, we distribute compute.
So all the devices that you put in, even if it's just a camera or if it's just a microphone, it has compute.
We can use that.
And then we need to bring new devices to the market with the latest generation chipset that are significantly more capable.
And when we do that, we can do more in terms of sort of reading the room and having more powerful models.
Typically still have to be sort of rather, narrow or kind of targeted.
But yeah, actually basically understanding more of what's going on.
So as this all evolves, we're also seeing other approaches to AI and hardware and AI on edge devices.
One example of this is there are a couple of different startups that are trying to do like AI pendants, for example, that are in some ways providing like a personalized meeting function.
What are your thoughts on these other approaches to AI hardware?
Obviously, they're serving related but slightly different audiences.
Just curious to get your take.
I think it's interesting.
What role does existing devices that you're used to have in this and to which extent is the room for new form factors and devices?
The jury is probably still out.
It tends to swing back and forth a bit.
Think, do you want it on the thing that's already on your meeting room or your laptop or your phone?
Or is there a room for something else?
What we know is that in the meeting room, people still need a device like ours because there's no way around needing a screen.
At least not anytime.
I think people are going to still want the screen for a fair amount of time.
They're definitely microphones.
They need cameras.
We're already there with those devices.
So for us, it makes sense to use that as sort of a platform or a system to deploy AI at the edge and in the shared spaces.
I think there's potentially some, and that's probably untapped, There's some but limited interaction between a personal device and a shared device in a space like ours.
So should our device, which is a shared device, to a larger or better extent leverage your laptop or your phone or whatever you're wearing in the future?
Probably.
I think that's why shouldn't it?
But yeah.
For what we're doing, it's not like there's this new magic form factor I think we'll put in the meeting room to solve the problems that we're solving.
There'll be more devices, more compute, new experiences, but some variation of form factors that you see today, probably.
More about seeing how you integrate with these other devices and how you can combine.
Yes, that's an interesting area.
Interesting.
Let's talk a bit about how Neat is using AI internally.
I'd love to understand your approach with your team, whether it's AI coding, where else you're leveraging it.
Obviously, AI is being integrated into your products, but it's always interesting to hear from leaders like yourself about how your team is actually adopting AI.
Yeah, that's another thing I spent quite a bit of time on.
I can imagine.
But I can't tell them, right?
So the thing here, I think, is to...
Just give...
Well, we have a lot of developers, obviously, that's where we see kind of the most significant use, like coding, AI coding, obviously.
But it's also used by hardware teams and marketing and other teams in the company.
But in general, just make sure that you give access to the tools that people want to try.
Sometimes it's good to not be the biggest company.
We're sort of a decent size now, but if...
If I or if the leaders want to make a tool available because there's something new that looks useful, we can do that in a few hours.
I can like Slack message.
Some assessment is needed sometimes, but we can move very quickly on making tools available.
So make sure that developers and the team has access to the tools that they need or that they want to try even to kind of become more effective.
Make sure that they are like encouraged to and have time.
to experiment and learn and do new things.
Because if you always feel that you're like incredibly busy, it can be kind of tempting just to revert to the old way of doing things.
That's not going to work now.
You just need to have time and backing to just like try something new, to try to do it in a new way, to try to use the tool to work in a different like repo or something else that you haven't worked on.
So like enough time and encouragement to experiment and learn.
Because this...
has to be driven to a large extent bottom up the individual like there's no way anybody can kind of sit and kind of tell people what to do or how to do it there's no course you can send people to there's no university you can go to to learn it you have to like just be you have to like want to experiment and learn and all the developers individually have to do that so that's like give them the tools give them like make sure they have like motivate like motivation they have themselves but like time backing to experiment and learn Make sure they learn from each other.
Because this is moving so quickly, some people will have learned a lot about this and some people will have learned a lot about that and some others would have failed by trying to do this.
So make sure you have the right forums or places where they can connect.
If it's like Lunch and Learns or demos or Slack channels or innovation days or whatever you want, connect, learn from each other.
And then I think increasingly it's also going to be about There's going to be some level of learning on structure and processes and roles.
Okay, so designer used to do this, product manager used to do that, developer used to do this.
Okay, but that doesn't necessarily have to be the case any longer.
How can we pull the design further into a simple implementation?
How can the product manager do a prototype and even start doing the feature rather than just doing the spec or doing the requirements?
How can a designer, how can a developer not have to ask for design because there's a skill that just does the design for that specific feature?
So like think about roles, think about processes, think about how you structure your repos and sort of making agents kind of more easily capable.
And then like be mindful of how and where.
you let the AI coding loose and it's going to be a bit different.
You're going to want to be more aggressive with sort of the adjacent stuff that's less dangerous and you're going to want to be a bit more careful with like your platform code or your DevOps environments and just like be mindful.
Be mindful and conscious decisions about where and how to use it.
So yeah, long answer.
Like some of the things we've been thinking about and working on.
It's a tough thing to grapple with.
It's fascinating to see how AI adoption is rapidly changing teams and how some are succeeding and some aren't.
But it's not just technical roles either.
Like I talked to a friend of mine in marketing earlier today and she was saying that this is the first company she's been at.
She recently took a new job where instead of having a CMS that she can make easy edits for, she has to now go through Git.
And that has become normalized because they're using agents to build the site and do most of the work on it.
And it's, you know, creating a learning curve for her and she's adapting.
But I think we're seeing this across every role.
Anyone who is driving forward with this is having to learn how they can integrate AI tooling into their own personal productivity stack for work.
And I think your point.
much earlier in this conversation about how NEAT is setting up a platform for people to build and leverage as far as having an MCP, the APIs that you're already independent with, but providing access to people's own meeting data so that they can leverage it in a variety of ways throughout their org speaks to the kind of approach that organizations need to use internally as well.
Okay, how can we open up opportunities for our team to build and succeed and how can we then provide the learnings that you pointed out of like both learnings from their co-workers but also externally hey here's what's happening so love that you're thinking through that and i think it's really crucial across every role today whether you are someone who's on the go-to-market side whether you are you know deep on the technical side uh or if you're in the executive layer you still need to be understanding what's going on too because To your point, it's not going to be 100% effective if you just say, oh, go token max.
So we can look at the complaints at Uber and Meta and elsewhere about token leaderboards and how they're suddenly having to pull those back because people have optimized for the metric in a way that is not necessarily healthy for the org.
Those tokens are not free.
So, yeah.
Yes, yes.
No, no, I think other functions like this.
To some extent, because this started with developers, they sort of know how to use these tools, right?
Because they're quite accustomed to kind of setting them up, configuring them, and they're built very much for their use case.
I think the other functions, there are some law, the simple stuff is law hanging, right?
You want to use it to do market research or help you with the presentation.
But if you want to use it for something that's like more core to your work, like if it's prospecting for sales or if it's marketing or finance.
somebody's going to have to help out a bit.
You can't just throw an MCP server and cloud the code at them.
There needs to be somebody there.
The companies who do that successfully will build capabilities in-house to partner as part of.
those functions or to partner with those functions to understand like their workflows, their needs, and how they can effectively deploy AI.
And that's going to need even more help than what developers needed because they're less technical functions.
Yeah, I had a great conversation with Angie Jones when she was at Block.
She's now at the Identic AI Foundation.
But when we talked, she talked about the rollout of Goose, their Blox.
AI agent to employees across the company and how it was fascinating to kind of go from this initial rollout to, oh, wait, I'm seeing, you know, non-technical team members use this coding agent to build their own MCPs to do custom tasks.
Like I've built quite a few MCPs for my podcast workflow, let alone like, and I want to say I've built them.
I mean, like I've asked Opus to help me build them typically, or I've asked Codex to help me build them.
So it's, it's fascinating to see.
how this is all changing and couldn't be more excited to see what happens with Neat as well in the next few years.
I think there is clearly a massive opportunity for you as you scale both the hardware and software side of what you're doing and then provide this open platform.
So would love just a closing thought from you about where you see Neat's future and for anyone who's interested in learning more about the company or checking out open roles where they should go.
Yeah, check out the website, Neat.no.
or reach out to me directly.
LinkedIn probably or wherever you can find me.
Closing, what we want to do.
So I love being in doing products when there are big changes and opportunities from a technology perspective.
I think that's fantastic.
I'm not here to just copy what everybody else is doing.
If you have new technology shifts like AI, it's like that's a very exciting time to...
to build products.
That's where we're at.
So we want to be at the forefront of making sure that the meeting room transforms into sort of a proactive participant in a meeting, that it bends into the background and just make sure that you can get on with, you make the meeting and we make it work, basically.
And we're going to be doing that in new and exciting ways with AI and other cool technologies.
I love that you bring up this idea of proactive participant because I think we're seeing that across technology right now.
And the meeting room is a huge example, but I think everyone is trying to make their tech stack more proactive with them.
And that's a lot of what we're doing with automation and with AI, with these non-deterministic models.
And it's a very exciting time, as you point out.
So couldn't be more thankful that you joined us for this conversation today.
It's been a ton of fun, Tormud.
I think our audience is going to love it.
So yeah, thank you so much for joining us.
It's been a pleasure.
Thank you.
Thanks for having me.
And for everyone listening, one last reminder, if you haven't subscribed on your platform of choice, whether that's YouTube, Spotify, Apple Podcasts, wherever you get your podcasts, please do.
And check out chainofthought.show and our newsletter at newsletter.chainofthought.show for so much more from the show, including transcripts, deep dives, takeaways, FAQs, and much more.
And we'll keep bringing you incredible deaths like Tormund every week.
Thank you all.
And Tormund, thanks again for joining us.