Thank you everyone for the opportunity to present here today. I'm very happy to be at the Atmosphere Conference and at the first Ad Proto Science Conference and tell you about my work. So I'm Sean Yumbluth. I'm a researcher at the San Francisco State Estuary Marine Science Center. I'm an adjunct professor there, and I'm I'm a former Berkeley lab transplant. I'm a biologist, oceanographer, and computational person, and I'm telling you about BioKea, this company that I founded and this first uh product that we're trying to build called Agentus. So what do we do? I'm working with a lot of insect uh enthusiasts, and we are running around California and we are sampling all the different parts and we are taking it back into the laboratory and my group we have robotics there where we're looking at these things in very, very high throughput, cataloging all the diversity of of life that's in California.
And then I'm leveraging the power of uh scientific storytelling through LLMs and trying to bring all this information to the light. And so BioKea, the company is the engine, and the product uh agentus is the output. So, what does a user journey look like here in the context of an atproto connection? Um the use case is imagining that you have researchers that are traveling to the Santa Monica Mountains to sample some dirt and then try to understand the scientific uh life that's inside of it. And so first uh the folks will ingest the data uh into our system, will analyze the data using canonical tools, we'll draft uh a manuscript, a very bare bones sort of data descriptor manuscript, and then we have uh also a reviewer engine built into this as well.
And then the final piece is to broadcast this information to the uh to the masses. And so this is uh just a quick mock-up of what the portal looks like right now. We're gonna sign in with our atproto account, giving everybody universal access through these decentralized identities. The first pay portal that you might get into is this like upload data type of portal where you're uploading your sequence data or images of plants or insects, anything that uh is kind of collected from from the scientists doing this sort of work. And then once you've uploaded your data, which it could be you know plants, uh DNA sequence data and and the like, uh metadata that that helps, we grind that through our analysis workflows and essentially give you a readout of the uh diversity information inside of your samples.
From there, because this is not ultra complex science, I feel confident in taking uh the use of AI to create some uh draft narratives from these things. And so you can imagine this rapid uh uh uh dissemination of work through through LLMs, where instead of having data products that are essentially waiting for analysis, we have a rapid ability to essentially communicate this in language that humans uh tend to agree upon. Alright, so once the papers are published, uh what does this look like? This is just kind of a glimpse into our current uh imaginative process here.
But we think that AI reviewers and human human reviewers are both going to be required. We know AI is going to become you know it's already a commodity, and so we can review papers with the push of a button essentially, but I do truly think we will always need humans in the loop who know what these things are and know what they're doing and the different categories here, for example, you know, novelty, we probably want that scored by a human more so than a robot. Um just kind of a glimpse into how we're approaching this. And so what do you do with with all this information?
So I'm an editor at two scientific journals and it's totally broken, and we can drink beers and pour one out for that broken system. Um and so the next system is uh this more public one. I think that could be built on AT Protocol and one that shows reviews in a transparent format. You know, we can share these things, they're immutable, they're reproducible, and they can be cited themselves. And then further um uh researchers they can start to take credit for these things. And this is a missing part of the scientific discourse that uh has basically completely been missed.
And so the science that I do, the the folks that uh collect these samples, they tend to do these like fun excursions where they go to like the California coast, for example. And so um what I'm imagining for uh something akin to like the Chai V prints or my version of that, are these uh interactive story maps that can be automatically generated from these data sets. And these are these nice digital artifacts that folks can explore and they're they're interactive and they're meant to help tell the story of this science. And then from there, what uh what can we do?
Well, we can hopefully amplify the stuff uh through platforms like Bluesky. And there's a bunch of different kind of teasers here for how this can all be connected, some including getting ahead of the uh point here, but um you can advertise your review that you completed it on Bluesky. You can highlight uh through Hypegen, which we haven't talked about here, but uh Joelle uh Chan was talking about that in a previous meeting that I had with him, and um, we have automatic uh hypothesis generation that can be captured. And then uh also a hat tip to symbol, which I love.
Um I'm imagining this could be a way that folks can build communities and try to gather trails of evidence that folks can then pursue uh with downstream work. Further, Matt. Uh so it it's uh very clear that the uh discourse graphs can play a huge role here, right? We we we all do science and we all appreciate how science is conducted, and uh this is the framework that I think can be very useful for uh for naturalists because um, for example, there's a whole lot of ancestral knowledge and like local knowledge that doesn't exist online, and I think that uh combination of discourse graphs and symbol and the rails that atproto represents can really capture this.
And so we're leaving the era of traditional publishing. We're no longer gonna be looking at static PDFs and Word docs, everything's gonna be more universal. We're gonna have submissions that accept raw data itself instead of your your uh written publics projects. Um the AI process is going to be faster. It's it's painfully slow right now. It's also going to uh continue to uh leverage human experts. Uh the output format, um, you know, dead paywall text is is uh old out with it. We're we are more excited about interactive story maps, knowledge graphs that are amenable to large scale ML applications, uh certainly the future.
And then on terms of in terms of data provenance, um clearly we have a much more connected system here where the decentralized uh app protocol can help link all these things together and folks can get credit for your reviews, your data production, your you know everything that we we agree is required for science. All right, so last slide. So the future ahead, and I don't know where we jump in this loop, but let's pretend we're we're starting at the top here. Um we're you know a bunch of bug nerds, we're running around California, we're collecting butterflies and all these things, and let's say we take those back into the laboratory and we look at the DNA and we take pictures of these things, and then we uh get to the bottom part here where we publish them with these sort of little interactive story maps, and then from there we can have these automated hooks to produce hypotheses from these data sets that folks uh online and and through uh connections with Semble, we can eventually get to a point at which you would say, aha, time for field campaign 2.0.
We're going to pursue this next set of hypotheses. And um in total, I guess I feel that this is the scientific loop that is possible if we build this and with the community that we're hoping to see.
Any questions for Sean? Thank you, Sean.
Random question. Uh, are you using the IGSNs at all for the inner the geosample numbers identifiers in what you're doing? Just curious. Yeah. Yes. Cool. Awesome. Yep. Trying to connect in all of the relevant um you know keys from everybody's databases. Nice. Okay. More questions. Thank you for shouting out Discourse Graphs. So what kind of schema do you think helps what kind of schema do you think helps put guardrails on the generation of the initial manuscript from the data. Hmm. I guess I want to understand. Do you do you feel like there will be too many folks that are able to contribute, or I guess I'm just trying to understand your like going from the data analysis to the scientific story, like maybe that's where you can plug in discourse graphs as well and do this this claim-based uh uh review the same way that like QED, this other AI uh review uh program does.
Yeah. Keep it if you keep the claim structure and each line of evidence as something that people can look back to that gets regenerated into these different narratives, then maybe it'll be more auditable. Perfect. Yeah, I think integration at multiple points with the the tools that you're building up is the way to go.
All right, let's thank Sean again and uh yeah, Alex. Feel free to come up.