Latest / Satellite image deep learning / Ai Agents with Hyades & BlazerOps

Transcript

Ai Agents with Hyades & BlazerOps · Satellite image deep learning

• 0:00
  [music]
• 0:05
  SPEAKER_00: Hi, Ash.
• 0:06
  SPEAKER_00: Hi, Jim, and welcome to the podcast.
• 0:08
  SPEAKER_01: Hi, Robin.
• 0:09
  SPEAKER_01: Nice to meet you.
• 0:11
  SPEAKER_00: Brilliant.
• 0:11
  SPEAKER_00: Let's begin with yourself, Ash, and a brief intro.
• 0:15
  SPEAKER_02: Sure.
• 0:15
  SPEAKER_02: Hi, I'm Ash.
• 0:17
  SPEAKER_02: I am one of the co-founders of Hyades and also the CEO here.
• 0:22
  SPEAKER_02: I come from a physics and computer science background where I worked
• 0:28
  SPEAKER_02: in developing algorithms to interpret and process data.
• 0:32
  SPEAKER_02: And we came across the complexities of dealing with various types of geospatial data
• 0:38
  SPEAKER_02: and have worked in that to be able to get to where we are with Hyades today.
• 0:44
  SPEAKER_01: My name is Jima,
• 0:45
  SPEAKER_01: one of the co-founders as well,
• 0:48
  SPEAKER_01: and I'm heading more so the partnerships with our customers as well
• 0:52
  SPEAKER_01: as leading operations.
• 0:54
  SPEAKER_00: Fantastic.
• 0:55
  SPEAKER_00: So Ash, what have you built?
• 0:58
  SPEAKER_02: Sure.
• 0:59
  SPEAKER_02: We have built Blazor Ops,
• 1:01
  SPEAKER_02: which is a platform to build complex machine learning models to interpret
• 1:08
  SPEAKER_02: and understand spatial data.
• 1:11
  SPEAKER_02: So when you have geospatial You obviously have an objective or an intent
• 1:17
  SPEAKER_02: that you are trying to understand from that data.
• 1:20
  SPEAKER_02: So obviously people in the geospatial industry build machine learning models to able
• 1:27
  SPEAKER_02: to interpret these kind of data.
• 1:29
  SPEAKER_02: So that process is really complicated and a complex process.
• 1:33
  SPEAKER_02: And we are streamlining that process by actually delegating those tasks to AI agents
• 1:39
  SPEAKER_02: to be able to build that machine learning models and the pipelines that you require
• 1:44
  SPEAKER_02: to build that machine learning models so that you can accelerate that process.
• 1:50
  SPEAKER_02: Cool.
• 1:51
  SPEAKER_00: And Jumin,
• 1:51
  SPEAKER_00: you're basically building something that is not targeted necessarily
• 1:55
  SPEAKER_00: at traditional machine learning people, but more generally.
• 1:57
  SPEAKER_00: Is that right?
• 1:59
  SPEAKER_01: Yes, that's right.
• 2:00
  SPEAKER_01: So machine learning engineers can use the platform as a power tool
• 2:05
  SPEAKER_01: to accelerate their processes,
• 2:07
  SPEAKER_01: but we ultimately envision this being a really helpful tool to unlock capability
• 2:11
  SPEAKER_01: for analysts and people who want to understand and get more in-depth nuances
• 2:20
  SPEAKER_01: and insights from their spatial data,
• 2:22
  SPEAKER_01: but they don't have the capability to create machine learning models to do so.
• 2:26
  SPEAKER_01: So we're hoping that this platform down the line will be able to unlock
• 2:29
  SPEAKER_01: that capability for analysts.
• 2:32
  SPEAKER_01: So the dream is really to connect your spatial data to the platform.
• 2:36
  SPEAKER_01: And with a plain English prompt, you can actually describe all of it.
• 2:40
  SPEAKER_01: And these agents will actually break that very big prop down
• 2:44
  SPEAKER_01: into these bite-sized actionable chunks,
• 2:46
  SPEAKER_01: run them in these compute spaces in parallel until it can arrive at
• 2:51
  SPEAKER_01: the most optimal machine learning architecture that's tailored to your objective
• 2:57
  SPEAKER_01: as well as your spatial data.
• 2:59
  SPEAKER_00: Brilliant.
• 2:59
  SPEAKER_00: And we've recorded a demo just to showcase those functionalities.
• 3:03
  SPEAKER_00: Now, Ash,
• 3:04
  SPEAKER_00: I'd really love to understand the difference between agents' workflows and sort
• 3:08
  SPEAKER_00: of more familiar LLM-based copy-paste kind
• 3:11
  SPEAKER_00: of workflows people might already have hands-on experience with.
• 3:15
  SPEAKER_00: Can you dive into that topic a bit?
• 3:18
  SPEAKER_02: Sure.
• 3:18
  SPEAKER_02: There exist agents in the LLM space and also in many other spaces right now.
• 3:25
  SPEAKER_02: For example,
• 3:26
  SPEAKER_02: when you have a coding agent that is assisting you with coding tasks in VS Code
• 3:32
  SPEAKER_02: or conventional ideas,
• 3:34
  SPEAKER_02: what it's really doing,
• 3:35
  SPEAKER_02: it's taking the ownership of the actual problem you have and then breaking that down
• 3:41
  SPEAKER_02: into multiple tasks and then Start calling or delegating that individual task into
• 3:50
  SPEAKER_02: an LLM model and then getting that output and then concordinating those outputs
• 3:55
  SPEAKER_02: and actually giving you back the exact structured output or the product or
• 4:00
  SPEAKER_02: the outcome that you want to be building using a coding tool.
• 4:05
  SPEAKER_02: So we can actually compare the spatial agents in a very similar way where
• 4:11
  SPEAKER_02: the problem then becomes, hey,
• 4:13
  SPEAKER_02: your spatial agents have to understand the data and then delegate tasks
• 4:17
  SPEAKER_02: to multiple either LLM or coding agents or any other kind
• 4:22
  SPEAKER_02: of building tools that's available in the bucket to be able to work
• 4:26
  SPEAKER_02: with these spatial data.
• 4:27
  SPEAKER_02: So very similar to LLM agents or coding agents,
• 4:33
  SPEAKER_02: spatial agents,
• 4:34
  SPEAKER_02: understands the data,
• 4:36
  SPEAKER_02: delegates tasks to various coding and other sort of models,
• 4:40
  SPEAKER_02: and it can coordinate all those evidences back in and then build that thing actually
• 4:44
  SPEAKER_02: for you.
• 4:46
  SPEAKER_00: So the sort of distinction between,
• 4:47
  SPEAKER_00: I guess,
• 4:48
  SPEAKER_00: the sort of stock coding agent and the geospatial agent is the tools that it has,
• 4:52
  SPEAKER_00: some sort of understanding, et cetera?
• 4:55
  SPEAKER_00: Yeah, yeah, that's correct.
• 4:57
  SPEAKER_01: So to provide more,
• 4:59
  SPEAKER_01: I guess,
• 5:00
  SPEAKER_01: context, we do use in the tech stack, regular LLM is provided by the Frontier Labs.
• 5:07
  SPEAKER_01: And so they are the ones that are ultimately doing the coding.
• 5:10
  SPEAKER_01: But you can think of Blazor Ops So,
• 5:12
  SPEAKER_01: I guess we've created that structure for the AI agents to break down
• 5:17
  SPEAKER_01: the complexities of a large machine-learning,
• 5:19
  SPEAKER_01: spatial machine-learning task into these bite-sized sub-tasks
• 5:22
  SPEAKER_01: that executive agents are then tasked to do.
• 5:23
  SPEAKER_01: And so
• 5:40
  SPEAKER_01: In a way,
• 5:41
  SPEAKER_01: we like to think about a platform as a harness platform that's tackling some of
• 5:45
  SPEAKER_01: the hardest challenges.
• 5:49
  SPEAKER_00: And you can also have multiple agents coordinating together.
• 5:52
  SPEAKER_00: I'd love to understand how that works.
• 5:57
  SPEAKER_02: Yeah, sure.
• 6:00
  SPEAKER_02: In this system,
• 6:01
  SPEAKER_02: you essentially have like an orchestrator agent who first takes up the data
• 6:07
  SPEAKER_02: and also takes your objective to understand what you really have to be developing
• 6:14
  SPEAKER_02: from this thing.
• 6:14
  SPEAKER_02: And that could be an LLM model.
• 6:17
  SPEAKER_02: Okay, and then you have other coding agents essentially like Cloud Code, Fable,
• 6:23
  SPEAKER_02: Opus 4.5,
• 6:24
  SPEAKER_02: even open source model like Kimi 2.5,
• 6:26
  SPEAKER_02: whatever is available in the bucket,
• 6:29
  SPEAKER_02: it can then Generate a prompt for a coding agent to create this preprocessing module
• 6:36
  SPEAKER_02: that I want it to be pre-processing the specific data.
• 6:40
  SPEAKER_02: Now,
• 6:41
  SPEAKER_02: for an analyst or an ML engineer or a data engineer who's actually doing this work,
• 6:47
  SPEAKER_02: have the capability to understand the requirements and then efficiently maybe prompt
• 6:51
  SPEAKER_02: a model or develop codes for this.
• 6:54
  SPEAKER_02: And the hardest part in this is really understanding the data because there exists
• 6:59
  SPEAKER_02: a visual aspect to this.
• 7:01
  SPEAKER_02: And unlike computer vision models,
• 7:03
  SPEAKER_02: which has been used conventionally for many tasks,
• 7:07
  SPEAKER_02: geospatial data is not that easy to interpret.
• 7:10
  SPEAKER_02: It comes with multiple layers of dimensionality, complexities associated with those things.
• 7:15
  SPEAKER_02: So we have actually developed models that are capable
• 7:20
  SPEAKER_02: of interpreting those multidimensional complicated data and
• 7:24
  SPEAKER_02: then efficiently give some feedback back to the orchestrator so
• 7:28
  SPEAKER_02: that they all can be informed about what really the objective or what is the state
• 7:33
  SPEAKER_02: of the model images are in terms of developing those codes and models.
• 7:38
  SPEAKER_00: I'd just like to really understand and clarify the difference in the models
• 7:41
  SPEAKER_00: that you developed,
• 7:42
  SPEAKER_00: as you say, and the LLMs that you're accessing that are from other providers.
• 7:47
  SPEAKER_00: Could you just provide some clarity on that?
• 7:52
  SPEAKER_01: Sure thing.
• 7:52
  SPEAKER_01: So in terms of the bespoke models that we've created,
• 7:57
  SPEAKER_01: one of the models that we created is our in-house foundational model pre-trained on
• 8:02
  SPEAKER_01: a corpus of hyperspectral data.
• 8:04
  SPEAKER_01: But that model essentially is able to ingest and provide inference for any kind
• 8:10
  SPEAKER_01: of hyperspectral, multispectral, and even just regular optical data.
• 8:14
  SPEAKER_01: So it's a model that can ingest any kind of spectral data full stop.
• 8:18
  SPEAKER_01: And the benefit of this is that we've actually embedded this within
• 8:22
  SPEAKER_01: a broader perception module within BlazeRops as one of the many tools that
• 8:27
  SPEAKER_01: the agents can use to pick up, to inspect and interrogate spatial data.
• 8:32
  SPEAKER_01: So when it realizes that it's dealing with spatial data and it wants to actually...
• 8:39
  SPEAKER_01: I guess,
• 8:39
  SPEAKER_01: tear apart the features within the spectral data into these dimensional clusters
• 8:44
  SPEAKER_01: to be able to better understand what different architectures to apply
• 8:48
  SPEAKER_01: to these different clusters.
• 8:50
  SPEAKER_01: It's able to actually pick up the foundational model as one of the many tools
• 8:54
  SPEAKER_01: to actually probe the spectral data.
• 8:57
  SPEAKER_01: And then based on that,
• 8:58
  SPEAKER_01: it's able to apply more nuanced architectural choices
• 9:02
  SPEAKER_01: and actually form more nuanced hypotheses for each kind of
• 9:11
  SPEAKER_00: Really interesting topic.
• 9:12
  SPEAKER_00: So do you anticipate that these agents will have access to multiple foundational models?
• 9:17
  SPEAKER_00: Do you think there's benefits to that versus having one sort of super model
• 9:21
  SPEAKER_00: that they all just use?
• 9:23
  SPEAKER_02: That's a really interesting concept because in geospatial,
• 9:28
  SPEAKER_02: it's almost impossible to build a supermodel
• 9:31
  SPEAKER_02: because all models are producing either classification or segmentation or
• 9:38
  SPEAKER_02: a desired probabilistic output.
• 9:40
  SPEAKER_02: Based on one specific intent, right?
• 9:43
  SPEAKER_02: So if you can do lots of clever engineering and R&D development like we did,
• 9:48
  SPEAKER_02: you know,
• 9:49
  SPEAKER_02: to accommodate all these types of modality ingestion to the data or the model.
• 9:55
  SPEAKER_02: You again are fundamentally bottlenecked by what you're going to produce, right?
• 10:01
  SPEAKER_02: Are you going to classify the data or are you going to predict the tree height
• 10:05
  SPEAKER_02: or are you going to predict maybe the picking up the deposits of some minerals
• 10:12
  SPEAKER_02: or also sort of things?
• 10:14
  SPEAKER_02: So they are individual models.
• 10:17
  SPEAKER_02: Now, all you can maybe do is to put all these models into a bucket
• 10:21
  SPEAKER_02: and give one endpoint like how Gemini 3.1 work.
• 10:24
  SPEAKER_02: It's really a multi-model mixture of expert model system where you can give a prompt
• 10:30
  SPEAKER_02: as some agent will pick up the right model to inference that specific data and
• 10:34
  SPEAKER_02: then give back to you the result.
• 10:36
  SPEAKER_01: And down the line as well, we also...
• 10:40
  SPEAKER_01: We'll be building out a feature so that if you've already built out
• 10:43
  SPEAKER_01: a really good proprietary model of your own,
• 10:46
  SPEAKER_01: then you want to actually Put into the toolkit for your agents to pick up, to use,
• 10:51
  SPEAKER_01: to actually probe and inspect the properties of your data, then you can also do so.
• 10:56
  SPEAKER_01: Currently,
• 10:56
  SPEAKER_01: the foundation model is one that we've already pre-trained,
• 10:59
  SPEAKER_01: we've developed, we know it works.
• 11:01
  SPEAKER_01: And so we've offered it as one of the tools for the AI agents to pick up when in
• 11:05
  SPEAKER_01: the inspection phase of the process.
• 11:08
  SPEAKER_01: But as you mentioned,
• 11:09
  SPEAKER_01: down the line,
• 11:10
  SPEAKER_01: we do want to open it up so
• 11:12
  SPEAKER_01: that users can actually upload their own proprietary modules for the agents to use.
• 11:19
  SPEAKER_02: We're also trying to incorporate as much open source weights out there
• 11:23
  SPEAKER_02: for these agents to access so that they can use those models as a base weight
• 11:30
  SPEAKER_02: to best find those models rather than building a model from scratch.
• 11:36
  SPEAKER_00: Yeah, and that sort of ties in with the sort of day-to-day work that I'm familiar with.
• 11:40
  SPEAKER_00: Do you have any kind of suggestions of if people want
• 11:43
  SPEAKER_00: to make sure their models are picked up and used by these systems,
• 11:46
  SPEAKER_00: what's best practice, et cetera?
• 11:50
  SPEAKER_02: Yeah, so especially I think Google Alpha Earth Foundation model is a good example.
• 11:56
  SPEAKER_02: If you are a model or analyst in a world bank and want to predict what's
• 12:04
  SPEAKER_02: the 2030 atmospheric temperature,
• 12:06
  SPEAKER_02: you know,
• 12:07
  SPEAKER_02: world temperature is going to be,
• 12:08
  SPEAKER_02: I have to inform all these industrial and government regulatory bodies that,
• 12:14
  SPEAKER_02: you know, this is going to really affect how the world will change.
• 12:17
  SPEAKER_02: So, someone who really have like a world-level analysis objective can pick up
• 12:24
  SPEAKER_02: a foundational model like Google Alpha Earth because it was pre-trained on
• 12:29
  SPEAKER_02: a wide variety of spectral synthetic picture and RGB data and then you can bring
• 12:35
  SPEAKER_02: in your world temperature variation map into that and
• 12:40
  SPEAKER_02: then correlate all those features and then you can use that model to predict things.
• 12:45
  SPEAKER_00: Right.
• 12:46
  SPEAKER_00: As more models come online,
• 12:48
  SPEAKER_00: the capabilities will expand,
• 12:50
  SPEAKER_00: and the richness of what those models bring will just enhance the system overall.
• 12:55
  SPEAKER_02: Yeah, yeah.
• 12:56
  SPEAKER_02: So foundational models alone are not deriving any immediate output.
• 13:04
  SPEAKER_02: You use those as a baseline for various applications,
• 13:09
  SPEAKER_02: and also the embeddings it creates can also be used for
• 13:14
  SPEAKER_01: interpreting the data
• 13:15
  SPEAKER_01: as well.
• 13:16
  SPEAKER_01: And we definitely believe that foundation models are one tool in the toolkit,
• 13:21
  SPEAKER_01: but definitely are not the silver bullet.
• 13:24
  SPEAKER_01: And ultimately,
• 13:26
  SPEAKER_01: what we believe is that
• 13:28
  SPEAKER_01: the whole reason why foundation models have What I have value is
• 13:32
  SPEAKER_01: that these embeddings greatly compress the development time and complexity
• 13:37
  SPEAKER_01: and they are a good starting point and in many jobs good enough to get the job done.
• 13:42
  SPEAKER_01: But when it comes to more,
• 13:44
  SPEAKER_01: I guess,
• 13:44
  SPEAKER_01: nuanced tasks where you do demand And if you create higher accuracy
• 13:47
  SPEAKER_01: and high performance of those machine learning models,
• 13:50
  SPEAKER_01: then it's in those situations where you do still need
• 13:52
  SPEAKER_01: to create those bespoke machine learning models.
• 13:54
  SPEAKER_01: And that's where we believe that Blazor Ops will have an enduring niche
• 13:58
  SPEAKER_01: to actually rapidly create those more complex and bespoke machine learning models on
• 14:02
  SPEAKER_01: the, I guess, with the idiosyncrasies of the dataset that you're bringing.
• 14:07
  SPEAKER_01: And I guess kind of tying back to the discussion about foundation models as well,
• 14:12
  SPEAKER_01: If the spatial agents form these hypotheses and it seems that it's appropriate
• 14:17
  SPEAKER_01: to actually take a foundational model,
• 14:19
  SPEAKER_01: let's say an open source foundation model and fine-tune it to get
• 14:23
  SPEAKER_01: the best possible outcome, then it will do so.
• 14:26
  SPEAKER_01: Or it could even take our foundational model that we've developed
• 14:29
  SPEAKER_01: and actually fine-tune it maybe with the ground truth dataset that you're providing
• 14:33
  SPEAKER_01: and actually use that To actually build the desired output model.
• 14:39
  SPEAKER_01: So again, it's got a whole toolkit of different options that it can choose from.
• 14:45
  SPEAKER_01: And it'll always try to go for,
• 14:47
  SPEAKER_01: I guess,
• 14:48
  SPEAKER_01: the most appropriate choice depending on your research objective and the data.
• 14:53
  SPEAKER_01: So it won't always have to go for a very complex model.
• 14:56
  SPEAKER_01: But when it has to go for those more advanced architectures,
• 14:59
  SPEAKER_01: BlazorOps definitely has the capability to be able to do so.
• 15:04
  SPEAKER_00: And on that topic of how a model would even be selected,
• 15:07
  SPEAKER_00: can you give a bit of insight into how that's done in practice?
• 15:11
  SPEAKER_00: Is it something to do with the details and the prompt or the complexity of the data itself?
• 15:18
  SPEAKER_02: Yeah, so there is a knowledge bank within the platform,
• 15:21
  SPEAKER_02: which is equipped with state-of-the-art publishments and also proven
• 15:26
  SPEAKER_02: and verified methods of doing stuff and proven pre-trained ways that have been used
• 15:32
  SPEAKER_02: and fine-tuned for various applications that's coming out of in research.
• 15:37
  SPEAKER_02: Those informations would become,
• 15:39
  SPEAKER_02: and it is actually actively managed and updated by us,
• 15:45
  SPEAKER_02: it becomes a first-hand reference tool for agents to work on.
• 15:50
  SPEAKER_02: If you give Claude maybe like a similar research question,
• 15:53
  SPEAKER_02: it would,
• 15:54
  SPEAKER_02: or any other LLM,
• 15:56
  SPEAKER_02: it would try to search on web,
• 15:57
  SPEAKER_02: which is kind of And
• 16:04
  SPEAKER_02: then it could go off the rail when searching web,
• 16:08
  SPEAKER_02: or even if it's given a very dedicated,
• 16:12
  SPEAKER_02: well-published paper,
• 16:15
  SPEAKER_02: it could go a little bit off the track,
• 16:17
  SPEAKER_02: but if you give the data in a very Predictable,
• 16:21
  SPEAKER_02: followable way for AI LLMs to read and pick up information from,
• 16:27
  SPEAKER_02: then you have less drift within those decisions.
• 16:35
  SPEAKER_01: Forming that as like a knowledge layer that compounds over time.
• 16:38
  SPEAKER_01: So that is a knowledge layer that benefits all users of the platform.
• 16:43
  SPEAKER_01: But there's also an ontology or knowledge layer that grows every time you as
• 16:48
  SPEAKER_01: an individual user or organization uses Blazorops.
• 16:52
  SPEAKER_01: So every single time you're creating an experiment or machine learning model project
• 16:55
  SPEAKER_01: on Blazorops,
• 16:57
  SPEAKER_01: The AI agents are actually keeping all of those trajectories or those learnings
• 17:02
  SPEAKER_01: in its own separate knowledge base.
• 17:04
  SPEAKER_01: And so when you're maybe down the line,
• 17:06
  SPEAKER_01: maybe one year later,
• 17:07
  SPEAKER_01: you know,
• 17:07
  SPEAKER_01: another employee is creating a kind of parallel or similar project to what you did
• 17:11
  SPEAKER_01: The AI agents will actually not be starting from scratch
• 17:15
  SPEAKER_01: and won't be wasting computer tokens trying to do that.
• 17:18
  SPEAKER_01: It'll actually draw as a first-hand reference from the company-wide ontology.
• 17:22
  SPEAKER_01: It'll see the similarities and it'll actually be able to build from those best practices.
• 17:28
  SPEAKER_00: I was quite interested that you mentioned
• 17:29
  SPEAKER_00: that you somewhat will curate this knowledge base as well.
• 17:33
  SPEAKER_00: I wanted to briefly talk about the topic of where The field is going,
• 17:36
  SPEAKER_00: in particular, the implications for people that are sort of domain experts.
• 17:40
  SPEAKER_00: Do you see that they'll be sort of increasingly there to curate kind of data that is
• 17:46
  SPEAKER_00: then used by systems and sort of less involved in the sort of day-to-day and
• 17:51
  SPEAKER_00: the details, or do you see it going in some other direction?
• 17:55
  SPEAKER_01: Well, I think this is the kind of bet that we're making as a company,
• 17:59
  SPEAKER_01: which is that if you democratize access to creating spatial machine learning models,
• 18:05
  SPEAKER_01: then we believe that the pie dramatically increases.
• 18:08
  SPEAKER_01: So geospatial industry is,
• 18:10
  SPEAKER_01: I think,
• 18:10
  SPEAKER_01: really And it's really interesting in the sense that it's an industry
• 18:12
  SPEAKER_01: that supports many industries.
• 18:14
  SPEAKER_01: And so you've got geospatial or spatial activity happening in places where you wouldn't imagine.
• 18:20
  SPEAKER_01: So for example,
• 18:21
  SPEAKER_01: taking insurance as an example,
• 18:22
  SPEAKER_01: many people don't associate that with a lot of geospatial activity.
• 18:25
  SPEAKER_01: But increasingly,
• 18:27
  SPEAKER_01: you have dedicated climate and catastrophe risk modeling teams
• 18:30
  SPEAKER_01: that are employing more and more sophisticated spatial ML.
• 18:33
  SPEAKER_01: Now, if Blazorop really is able to democratize the development
• 18:38
  SPEAKER_01: of these spatial machine learning models,
• 18:40
  SPEAKER_01: Anticipate that there will actually be an increased need
• 18:43
  SPEAKER_01: for really sophisticated machine learning engineers.
• 18:47
  SPEAKER_01: And also we believe that It'll also mean that there'll be greater value placed
• 18:51
  SPEAKER_01: on machine learning engineers who can really push the boundaries of ML as a field
• 18:55
  SPEAKER_01: and,
• 18:56
  SPEAKER_01: as you mentioned, introduce new knowledge and curate new knowledge into the world.
• 19:00
  SPEAKER_01: And we believe that there will still be a space for really that kind
• 19:04
  SPEAKER_01: of like bleeding edge,
• 19:05
  SPEAKER_01: you know,
• 19:06
  SPEAKER_01: foundation model development and sophisticated ML technique development,
• 19:10
  SPEAKER_01: architectural development.
• 19:12
  SPEAKER_02: A lot of ML researchers are coming from a background to research and find new ways
• 19:18
  SPEAKER_02: to do things,
• 19:19
  SPEAKER_02: but they are always stuck in the loop of developing custom things that are required
• 19:24
  SPEAKER_02: for the immediate need of an analyst or a company.
• 19:27
  SPEAKER_02: So if we can get those tasks off their hands to blazer ops,
• 19:32
  SPEAKER_02: this can actually do those things for all jobs,
• 19:34
  SPEAKER_02: then the ML researchers and scientists can actually put their valuable time
• 19:39
  SPEAKER_02: in discovering and inventing new things that would benefit Wonderful.
• 19:46
  SPEAKER_00: Very exciting time to be alive.
• 19:48
  SPEAKER_00: I sort of share your optimism and hope that we're going to grow the pie
• 19:51
  SPEAKER_00: and bring lots more people to the geospatial domain and really expand and make
• 19:55
  SPEAKER_00: the maximum impact from the data and the skills that are there.
• 19:59
  SPEAKER_00: Let's round things up here.
• 20:00
  SPEAKER_00: So the website is hides.com?
• 20:05
  SPEAKER_02: The website is hides.space, but the platform is blazerops.ai.
• 20:11
  SPEAKER_00: Right, blazerops.io.
• 20:13
  SPEAKER_00: I'll put the link in the show notes.
• 20:15
  SPEAKER_00: I assume that's where you publish updates on your platform, et cetera?
• 20:20
  SPEAKER_02: Yes.
• 20:21
  SPEAKER_00: Wonderful.
• 20:21
  SPEAKER_00: I'll add that link as well.
• 20:23
  SPEAKER_00: Well, once again, gentlemen, thanks for joining me today.
• 20:25
  SPEAKER_00: Really interesting conversation and all the best with the launch.
• 20:28
  SPEAKER_01: Thank you very much.
• 20:29
  SPEAKER_01: Thanks, Robin.
• 20:29
  SPEAKER_01: Appreciate it.
• 20:31
  SPEAKER_00: Yeah.