WEBVTT 00:00.000 --> 00:15.860 So, hello everyone, this is also going to be a short presentation. 00:15.860 --> 00:21.240 Together with Mathias and Christina, we will present you a picoscolor, which has the aim 00:21.240 --> 00:25.560 to advance open research with an open source AI platform. 00:26.440 --> 00:31.920 Currently, we are an open source proof of concept, this is our stage, and we are trying to leverage 00:31.920 --> 00:37.000 domain adapt to the AI, and real-time collaboration to create this tool. 00:37.000 --> 00:42.520 And our goal in the future is to gather the community of developers and researchers 00:42.520 --> 00:47.360 around the tool in such a way that we can advance systematic reviews. 00:47.360 --> 00:52.920 We are a group of master and computer science students from Georgia Tech, our story started 00:52.920 --> 00:57.400 with, initially, a team of four software engineers, then we moved to five software engineers 00:57.400 --> 01:04.280 and we participated to two hackathons, and also we collaborated with Robert Gordon University 01:04.280 --> 01:12.720 research researchers, and with the help of DevPost, we actually managed to win a hackathon. 01:12.720 --> 01:17.920 We got second prize in the Red Hat and AI hackathon, which actually motivated us to continue 01:17.920 --> 01:25.120 with this project and try to move it from a proof of concept to an actual tool that 01:25.120 --> 01:28.160 people can use and be useful. 01:28.160 --> 01:33.520 The reason why we want to do that is because we believe that currently there is an existing 01:33.520 --> 01:40.480 challenge, a user's face certain issues, and existing tools in the field of systematic reviews 01:40.480 --> 01:47.480 struggle with increasing volume and complexity of technical literature, specialized terminology, 01:47.480 --> 01:53.000 transparent selection workflows, and also there is a limited ability to integrate new 01:53.000 --> 01:58.720 publications into ongoing or past literature reviews. 01:58.720 --> 02:03.640 In terms of the users that we believe that might find this tool useful, first of all, the 02:03.640 --> 02:08.680 academic researchers, we believe that it's challenging for them to navigate through vast amounts 02:08.680 --> 02:15.240 of literature and new knowledge is hard to integrate and publish literature reviews. 02:15.240 --> 02:18.920 So conference organizers, imagine that you have to organize like a conference with thousands 02:18.920 --> 02:24.800 of publications, and you have to scheme through all the articles that are being submitted 02:24.800 --> 02:31.400 to the conference, and this tool is actually helping to avoid manually reviewing submissions, 02:31.400 --> 02:39.480 and also avoid a situation that might lead to time-consuming and potentially bias selection 02:39.560 --> 02:43.560 processes, and also students and educators. 02:43.560 --> 02:49.960 They currently lack access to comprehensive resources and structured guidance for conducting 02:49.960 --> 02:53.640 effective literature reviews. 02:53.640 --> 03:00.520 In terms of how we solve this problem, this profile concept currently leverages AI powered search 03:00.520 --> 03:06.760 and ranking, which helps with enhancing literature reviews, literature relevance. 03:06.760 --> 03:14.040 So basically, when we search, it's not just like a typical search tool, it uses AI to search 03:14.040 --> 03:19.720 through all the papers in such a way that the search becomes more relevant. 03:19.720 --> 03:26.600 It has a scalable architecture, uses microservices for flexibility, and this in the future 03:26.600 --> 03:32.200 will help us adapt and use different, for example, plug-and-play models. 03:32.200 --> 03:38.360 So we are thinking about collaboration tools, and the tool is designed for shared work spaces 03:38.360 --> 03:45.480 and role-based access, and in terms of the systematic review flow that we have, this includes 03:45.480 --> 03:51.800 metadata extraction and selection criteria, and currently it's inspired from the biomedicine 03:51.800 --> 03:57.160 methodologies, actually peak of stance from population intervention, comparison, and outcome. 03:57.160 --> 04:02.600 And yeah, if you didn't understand how the tool works so far, I will do a little demo for you. 04:04.200 --> 04:08.120 This is one of the main functionalities, this is the search tool. 04:08.120 --> 04:11.880 So you can see here the user inserting the search. 04:11.880 --> 04:17.960 You can see that there's an advanced search option with also the possibility to have scientific 04:17.960 --> 04:24.920 notation translation, and also we have an AI query optimizer that helps users optimize their 04:25.320 --> 04:36.920 question for the search tool. After inserting the question, basically we get to this funnel section 04:36.920 --> 04:43.080 that has an AI summary on top with all the papers that are being displayed on the first page, 04:44.200 --> 04:50.840 and the user can select the paper and move it throughout the funnel and give it different 04:50.840 --> 04:57.000 other statuses. For example, in this case, it has moved from identified to screened and also 04:57.000 --> 05:01.480 the user can archive that selection and also bring it back, of course. 05:04.840 --> 05:12.280 And also it has the feature of showing only the items that has PDFs. For example, if we get 05:12.920 --> 05:18.120 in a search, we get like 2000 results, for example, but not all of them has half PDFs. 05:18.680 --> 05:26.440 This will just retrieve only the items with PDFs, and we have the possibility of chatting 05:26.440 --> 05:31.240 with a PDF AI assistant about that PDF in specific, and you can see a display here. 05:31.240 --> 05:35.480 Now let me tell you how to explain you more about the rest. 05:36.360 --> 05:50.760 Okay, hopefully that works all right. Hi, everyone. Thanks for your attention. 05:52.600 --> 05:57.160 So yes, so now we covered the problem with the space as well as some of the key pain points 05:58.280 --> 06:03.400 and you'll watch the tool in action. So now I want to show you how because 06:03.400 --> 06:09.400 color works behind the tools. So we behind the scenes and also what we see for the future 06:09.400 --> 06:14.040 because the work where are proof of concept, right, but we want to push this into like MPP, 06:14.040 --> 06:20.360 where it becomes functional for the users. So here's a high level view of a because color and 06:20.360 --> 06:25.640 the solution. Essentially, users have the ability to upload their own documents and this triggers 06:25.640 --> 06:31.960 a data cleaning pipeline. After that pipeline, we have a fine-tune AI models that will perform 06:31.960 --> 06:36.840 metadata extraction. So in the case of biomeric and we extracted the peak elements, but 06:36.840 --> 06:45.320 the idea is to tell all these specific domains with fine-tune adapters. So once we have this, 06:45.320 --> 06:50.440 we embed everything so the documents as well as the metadata and store it on a vector store 06:50.440 --> 06:56.840 that is able to perform semantic and similarity search. So yeah, in essence, like we're ranking 06:56.840 --> 07:04.040 these stories. Finally, we have a rack chatbot. So this can plug into like your favorite LLM 07:04.600 --> 07:10.360 and that can provide some rich summaries and insights. Additionally, if you see all of those 07:10.360 --> 07:15.400 like feedback loops, that's the automatic tension where every annotation or user decision 07:16.360 --> 07:21.560 fits back into the model in that making this whole process smarter and more accurate over time. 07:22.440 --> 07:30.360 In terms of our current stack, we use fast API and Python and also like power our AI features 07:30.360 --> 07:35.880 with fine-tune cyber and LLM index. We have re-acting the in the front end and we have 07:35.880 --> 07:42.440 containerized all of these applications. Well, we mentioned before, we participated in Red Hat 07:42.440 --> 07:47.880 and Intel Hackathon. So you know, we actually the hosties on the Red Hat OpenShift environment 07:48.040 --> 07:52.760 and we take a advantage of Intel OpenBeno model formats for faster inference. 07:53.560 --> 08:00.920 And this data, the moment sits on TIDB for scalable vector search. So I think that you have 08:00.920 --> 08:06.120 a free tier that we were able to use, but in the future we would like to host or 08:06.120 --> 08:09.880 provide the path to host the solution and full like open source infrastructure. 08:10.920 --> 08:14.440 We have microservices and we have AI adapters. 08:15.400 --> 08:20.680 So in terms of the potential and how we see this tool is like surrounding like some key aspects, 08:20.680 --> 08:26.440 right? So it's all about like this principle of collaboration and transparency within the systematic 08:26.440 --> 08:31.880 review and how you know let's say a professor can interact with their TAs and form their own projects. 08:31.880 --> 08:35.960 But as well, how can that move like a cross project, right? And how can we perform like 08:37.800 --> 08:42.200 archive of different annotations that could be trained for specific domains. 08:43.160 --> 08:48.840 For this human AI collaboration is quite important for us, as well as we aim to integrate 08:48.840 --> 08:55.000 with other tools such as automatic annotation tools or reference annotation managers. 08:56.280 --> 09:00.200 We want to create an intuitive user experience, hopefully you got that from having a look at the 09:00.200 --> 09:09.640 demo. So dashboard and data driven insights. So we are here on the POC and we aim to go into 09:09.640 --> 09:15.800 a self-host solution. We have some personas which I'll talk in the next slide, but essentially 09:15.800 --> 09:21.000 we're taking one little like thin slice, produce a valuable product there and then we plan to 09:21.000 --> 09:25.480 expand for a more complicated use case like, you know, full-on system attribute. 09:26.440 --> 09:31.160 So in terms of our first persona that we aim to achieve is the conference organizer. 09:31.160 --> 09:35.960 And we picked this because it's quite like a close-knit, bring your own documents type of a scenario. 09:36.840 --> 09:42.040 Currently, like in terms of this visual, you can see that we are nearly the end of the phase one, 09:42.040 --> 09:46.920 which is all about the back end and the core features. These are like the growth operations for 09:46.920 --> 09:53.640 management documents, the AI driven, driven as core in, so on and so forth. And on the next phase, 09:53.640 --> 09:57.880 we want to focus more on, you know, how to make this collaboration and more user management. 09:59.240 --> 10:04.040 And yeah, as well as increasing the efficiency of our ranking processes. 10:04.760 --> 10:10.760 Lastly, provide the PI2 on top of all of that. So this is just a little bit more of, you know, 10:10.760 --> 10:15.480 once we nailed that first scenario, our focus areas are just improving the system, improving the 10:15.480 --> 10:20.600 automation and connecting to more data stores, right? So at the moment on our demo, we only have, like, 10:20.600 --> 10:26.840 you know, 30,000 documents, but we want to link more to databases like both methods or archive. 10:27.480 --> 10:33.240 Turns off, you know, if you want to contribute, we welcome contributors, especially, 10:33.240 --> 10:38.200 like some back end developers, right? So if you like go, for instance, we're applying to to build 10:38.200 --> 10:43.640 a back end based on that, as well as you, if you make very intuitive interfaces, come and talk to us. 10:44.520 --> 10:49.480 Ultimately, you know, I think we're interested on developing a lot of the AI models, and we have 10:49.480 --> 10:54.840 some help from Intel and some of the processors to the hackathon. So if you're interested on those 10:54.920 --> 11:01.400 technologies come talk to us, or if you're a user, right? If you want to use this proof of concept, 11:01.400 --> 11:06.760 now for your current project or, you know, maybe on the future projects or to change these two 11:06.760 --> 11:12.920 towers, you know, something that might help you, we're happy to hear your ideas. In terms of how to 11:12.920 --> 11:19.080 join this community, so we have our GitHub repo, currently we're meeting weekly, so usually on 11:19.080 --> 11:26.440 two states, and yeah, we're happy to engage online as well. So think that's the end of the presentation. 11:49.720 --> 12:02.600 Sure, so the question was, what do we use for the extraction pipeline? I assume you 12:02.600 --> 12:08.520 mean the metadata extraction. Yeah, so that's how like the initial stage of this two, right? So 12:08.520 --> 12:14.120 it was based on one paper from Robert Gordon University on training adapters, specifically for 12:14.120 --> 12:23.160 biomedicine, and there are other open source, also data basis for extracting people. So we use 12:23.160 --> 12:29.160 that about like a thousand records, and I'm fine killed a bird model, a cyborg model, and that will 12:29.160 --> 12:36.040 become our metadata extractor. So at that point, you know, it knows how to extract these four key elements 12:37.240 --> 12:42.040 put it into our tool. And we do some forther cleaning on top of that, right? But the idea is like if you 12:42.600 --> 12:47.800 bring in other domain with the core elements for your systematic review of your selection criteria, 12:48.520 --> 12:54.680 you can decide those four, five key elements, bring in a thousand annotations, and then you can't 12:55.560 --> 12:58.040 explore the literature at the scale.