Listen on
Overview
The sheer, ever-growing volume of scientific information presents a critical bottleneck for biotech innovation. Research teams, business development, and investors in life sciences spend significant time sifting through patents, literature, and clinical trials to de-risk new ventures or secure intellectual property. Without a reliable way to distill this information, companies face delayed decisions, misinformed strategies, and missed opportunities for scientific breakthroughs.
Alexander Junge, Co-Founder and CTO of Amass, brings his bioinformatics expertise to this challenge. Amass developed a scientific intelligence platform that provides transparent, verifiable answers to complex scientific questions. This episode details Amass’s unique approach to integrating vast public and proprietary datasets, employing advanced information retrieval, and using AI to generate consistently reliable insights—a critical distinction from consumer-grade AI tools.
Ross Katz and Alexander discuss the technical strategies Amass uses to eliminate AI inaccuracies by grounding every output in source documents. Alexander shares a compelling case study: how Amass’s AI assistant partnered with Nordic Bioventures to accelerate biotech ideation, leading to the validated concept for FiiSH, a company focused on sustainable fish feed. The conversation also explores how professional-grade AI tools move beyond simple chat interfaces to support deep scientific inquiry and document generation, all while maintaining rigorous data security and privacy.
Key Takeaways
Prioritizing Rigor Over Speed is Essential for Scientific AI
In professional scientific contexts, the trustworthiness of AI-generated answers outweighs immediate response times. Amass designs its systems to ensure reliability through multi-stage processing, including generating consensus answers and cross-checking them against original source documents. This deliberate focus on accuracy, even if it means a minute-long processing time, builds confidence in the system’s outputs for high-stakes decisions.
AI Accelerates Early-Stage Biotech Ideation and Validation
AI can significantly compress the timeline for de-risking new biotech ventures. Amass’s platform assisted Nordic Bioventures in identifying promising ideas, assessing patent whitespace, and rapidly planning killer experiments for validation. This capability directly translates to faster company formation and quicker access to external funding, demonstrating a clear ROI for AI in venture creation.
Professional AI Tools Require Tailored Interaction Models Beyond Basic Chat
Generic chatbot interfaces fall short for complex scientific research. Amass has evolved its platform to offer ‘scientific deep dives’ where users define scope and select dimensions, generating detailed, long-form reports. Additionally, ‘document-first’ interactions support tasks like grant writing, allowing experts to refine and collaborate with the AI on specific sections, ensuring control and contextual accuracy.
Data Security and Trust are Foundational for Proprietary Scientific AI
When handling sensitive scientific data, strong security and privacy are non-negotiable. Amass hosts its solution on established cloud infrastructure like Microsoft Azure, building on existing client trust. Crucially, the platform avoids training its models on user-provided internal data, ensuring proprietary information remains confidential while still allowing the system to learn from inspired, anonymized queries.
Related: CorrDyn works deep in biotech and life sciences and provides data engineering services to build reliable systems. Learn how biotech manufacturers gain more from their data and our approach to AI strategy.
Full Transcript
Jason: Welcome to Data in Biotech, a podcast from CorrDyn where we explore how companies leverage data to drive innovation in life sciences. Every two weeks we sit down with an expert from the world of biotechnology to understand how they’re using data science to solve technical challenges, streamline operations, and further innovation in their business. In this episode we sit down with Alexander Junge, co-founder of Amass, to discuss his journey from a background in bioinformatics in Germany to co-founding a life sciences intelligence platform in Denmark. Alexander walks us through the Amass platform, which helps life science professionals get reliable, transparent answers to their scientific inquiries by analyzing scientific literature, patents, conference papers, and more. We touch on the technological challenges and solutions for data integration, search, and information retrieval, as well as how Amass ensures data accuracy and combats issues like hallucinations in AI outputs. Alexander also shares a case study involving Nordic BioVentures, highlighting the practical applications of Amass technologies in real-world biotech innovation. Here we go.
Ross Katz: Alexander Junge, welcome to the Data in Biotech podcast.
Alexander Junge: Thank you. It’s great to be here.
Ross Katz: Well, just to kick us off, would you give us a brief introduction to your background and what brought you here today?
Alexander Junge: My name is Alex, I’m originally from Germany but been living in Denmark for the last, I think around 12 years. Have a background in bioinformatics, worked in pharma at Novo Nordisk for some time, at Corti, a local health tech company here in Copenhagen, and then co-founded Amass about one and a half to two years ago together with my co-founder Henrik. He’s the commercial arm of the founding team, I’m the technical and scientific arm of the founding team.
Ross Katz: Cool. Would you give us an introduction to Amass, since we’ll be spending some time talking about your platform today?
Alexander Junge: Of course, I’d be happy to. What we’re building is a scientific intelligence platform—basically a platform where our users and customers can go to get answers to scientific questions in the life sciences. We are focusing on the life sciences domain and the types of questions our users ask is what are the medications treating a certain condition if they’re in the therapeutic space, if they’re working with engineered microbes, they might be interested in what are the common challenges people are facing in the wet lab, or it might also be questions regarding intellectual property—patents they should be aware of in their space of operation or the bio-space they’re trying to either do research in or build a company in. We are basically building a platform where the user can come in and get answers to these kinds of questions in a very transparent and reliable manner. What does that mean to be transparent and reliable? One is that we ground the answers in actual data and reliable data—biomedical papers, conference papers, publications, preprints, but also patents, clinical trials if that’s relevant, or also internal documents that organizations, especially large ones, have acquired over the years that contain a lot of reliable scientific information. Then what the platform does, when the user comes in with a question, we’re looking across all these different sources, saving a lot of time, looking into way more resources than any human user would, and providing a reliable answer that is very thoroughly sourced and transparent so the user can figure out where it comes from and go deeper, asking additional follow-up questions.
Ross Katz: You come to this from a bioinformatics background. You have a PhD in bioinformatics and you’ve spent time in pharma. I’m interested in how does this background and these experiences shape your understanding of the nature of search, information retrieval, research in life sciences today?
Alexander Junge: One of the projects I worked on during my postdoc in bioinformatics was a protein-protein interaction database called String. It’s actually used by tens of thousands of researchers around the globe every week. That gave me an interest in building products that actually enable wet-lab researchers to look into way more data sources than they were ever able to look at. I was only a small part of the String database team, but I was able to contribute to that and see what impact one can have building tools for researchers—in this case enabling them to analyze their protein-protein interaction data coming out of the wet lab. Huge datasets, researchers have a really hard time making sense out of this data, and we basically built a tool that integrated information from a range of resources—experimental data, but also the literature. Reading through papers, or using computers to do that and extracting relevant information from that. That got me a taste for, when I was working on this about 10 years ago, that there is really something to be done here and impact to be created. Back then the technologies we were using were somewhat primitive by today’s standards. I remember trying to have neural networks and transformers generate text 10 years ago—stuff that we take for granted these days didn’t really work back then. Since then with the whole transformer wave and generative AI, scaling up the technologies we were looking at back then, it’s very much possible to build very interesting product experiences for end users right now. I also saw the need for building tools like we’re doing at Amass when I was working at Novo Nordisk—I could see that very large companies have these kinds of problems too. The question-answering is the base foundation we have in Amass on our platform right now, but we actually have a range of additional tools and user experiences we’re building on top of that, where our AI-based research assistant called Gemma helps the user figure out what it is exactly they could be asking in the first place. Sometimes you’re looking at a problem space where you’re not the expert in a given domain, so we’re also building experiences that help our users figure out what it is they’re really after.
Ross Katz: Yeah, you just named one of the core challenges of scientific research, which is knowing what question to ask and the way to formulate that question so that the responses are actually answering the core question they’re trying to get at. Is that what you would consider to be most challenging about scientific research today, or are there other challenges that you’re really setting out to solve with Amass?
Alexander Junge: That’s the primary challenge we are tackling with Amass. Of course there are all kinds of other challenges—reproducibility is a common challenge in the sciences. Not so much something we’re focusing on right now, that’s what others are doing. A lot of that touches on the wet lab, etc., but we are really focusing on the scientific information retrieval problems—the fact that science is, you’re perceiving your work as standing on the shoulders of giants. But the problem is these giants, they grow every week or every day or every year to even larger sizes. The sheer amount of information that you need to know and study throughout your academic career as a young researcher just keeps growing over time. We’re essentially building tools to help with this scientific information challenge in the life sciences.
Ross Katz: Yeah, that makes a lot of sense. At the beginning you talked about the breadth of datasets available—medications treating certain conditions, engineering microbes, intellectual property, patents, biomedical data, conference papers, clinical trials. The list just goes on and on. But then what you just described was the depth of knowledge—over time in each of these domains there’s just more and more data accumulating. Can you talk about how you approach integrating knowledge across both the breadth and the depth of things people might be asking questions about?
Alexander Junge: It’s interesting that you say that you can both go broad and also deep on specific sources. What we have internally is multiple big databases of literature and patents and trials and additional sources where a lot of the information is actually public, but it’s just very hard to find in the public domain. We make that searchable and findable in our own systems and we need to do that in a flexible manner because the kinds of questions we get on the platform come from a wide range of biomedical or biological domains. When we first take the data into our platform, we don’t need to cater to one specific use case—we keep it flexible and then focus on when the user asks a question, translating those searches into the way we have indexed the data internally.
Ross Katz: When you’re dealing with the breadth of document types that you’re talking about, the different types of questions that might be asked are equally broad. There are so many different roles that people might have, from wet-lab researchers to IP lawyers asking questions of your data. Can you characterize the different jobs-to-be-done or the different roles that you see interacting with Amass?
Alexander Junge: In the grand scheme of things, I would say the class of jobs-to-be-done for us is whenever our users or customers need to venture into a scientific domain, a life sciences domain, or a question that they’re not the masters in already. It’s not really that we’re catering to a PhD-level student who should very well know their research topic and all the key papers in it. But if you look at how innovation teams in large organizations work, or business development teams, or smaller biotechs—which is actually our largest customer base right now—or even life sciences investors, they come across a lot of scientific questions every day where they need to make decisions quickly under uncertainty. At the end of the day, the class of problems we’re trying to solve is living in that space and removing some of that uncertainty by producing reliable answers to the kinds of questions we’re getting. What is known about a certain topic, what should I be very much aware of when venturing into this new space.
Ross Katz: Yeah, that makes a lot of sense. I know that you were part of one of the first cases where an AI-assisted venture organization helped do some ideation and create or shape a biotech company. Can you share more about that use case and how it worked?
Alexander Junge: I’d be happy to, because it nicely illustrates the value our platform and products can create in the real world. What we did there was work with a local venture firm here in Copenhagen, Nordic BioVentures. What they’re doing is essentially evaluating a lot of ideas for new biotechnology companies all the time, and they were working with our life sciences research assistant Gemma. That happened last summer, summer 2024—a few months ago. Since then we’ve progressed our technology. But what we were able to do back then was assist them in what would be, in a design thinking world, a classical double diamond process. They had different ideas sketched out on whiteboards—could we be doing X or solving Y and starting a new company around this. We fed this into our AI research assistant Gemma to first figure out, among all these diverging ideas, what are the most promising ones to look into. That’s a multifaceted, multidimensional question. Of course you want to know what’s the IP space, patents are super important for biotechnology companies especially early on. Is there white space in the patent landscape? What does the research say—is this a moonshot or is it a realistic endeavor to achieve in a foreseeable amount of time? The experts in the venture creation studio worked hand-in-hand with our assistant to identify the most promising ideas. That resulted in around three or four ideas they went forward with. Then they went further—the second diamond is again diverging, after picking the problem for the company to work on, on how to actually go about it. How do you find a solution? In this case it was a lot about planning experiments, coming up with killer experiments to validate the idea as quickly as possible. Our assistant was able to very quickly help them understand this new domain, plan those experiments, figure out what kinds of experimental setups they needed to zoom in on—this is how we do it, this is how we validate the idea. The first idea that was actually validated is a company called FiiSH. We’re talking about this because the company has now gotten its first external funding—one of the first big milestones you need to reach when starting a biotech company. They’re working in the space of producing sustainable fish feed by fermenting insects. Gemma was working more as a partner in this whole venture creation. It’s an interesting case because it shows what might be possible with these assistants in the future. Maybe the last thing I want to say here is that this is the one company we can talk about publicly that was created, but there’s actually another one going through the last experimental trials right now—so we can show this is not something that only happened once, and we have first indications.
Ross Katz: I’m interested in—Nordic BioVentures I’m assuming has their own framework for what questions they ask in order to ideate about these companies, and they also have a framework they apply to evaluate the outputs of those questions and determine which are the highest priority spaces and solutions. Can you give me an idea of what the venture organization brings to the table and what Gemma brings in terms of helping with the refinement?
Alexander Junge: The process went—as you mentioned, they have almost like a questionnaire or at least a basic list of dimensions they want to be looking at, not super concrete questions. Gemma first helped to shape what are the concrete questions we should be asking for this one specific company. What does it mean for troubleshooting experiments when you’re working on a more sustainable fish feed alternative? Very specifically then formulated questions around that—which insects, for example, should you be fermenting to reach the optimal nutritional profile. Then zooming in on these different dimensions and specific questions together with the user. What the people in the venture studio brought was their expertise. We don’t have access to all the data in the world—it’s very transparent what we’re integrating in our platform—we don’t have the latest market analyses that they might know about or bring in. A lot of the role in this interaction between AI and expert was the expert reading the answers from the AI, double-checking it, evaluating what is the most promising route among different options suggested by our tools, then zooming in on any one of them. It was really a partnership. There’s research on this where people did actual studies with companies and I think what comes out of that research is that the people who benefit most from AI-based partners are actually the ones who have a lot of expertise in the domain already. That’s also a pattern we saw here. They knew what they were doing and our assistant was a partner in making that process more scalable and much faster.
Ross Katz: Yeah, that’s interesting. Earlier you mentioned that the organizations that interact with Amass have a repository of data you’ve already indexed, and the party interacting with it also brings their own data in. But the interaction pattern you just described is them checking the work of the agent against research they had on hand. Is there a decision about what information they expose to Gemma or to Amass versus what they keep out? And how do the decisions about how to configure the system influence the outputs they would get?
Alexander Junge: Taking the first question first: what was the data that went in? It was public data in this case. All the data types I explained earlier—trials were not super important here—but patents and research papers from all kinds of domains, not only biomedicine. That was purely external information that you would be able to find yourself given infinite time, and no one has that. The second question is how much internal secret sauce data we needed. Again, for these examples, we didn’t really use any. There’s some expertise in the kinds of questions the user is asking to work with the AI in partnership. One of the things we are not doing is training on any user data. The data that our users put into the system—internal documents or the questions they’re asking—it stays with them and we’re not training new models to then show it to other customers. That’s super sensitive data, but it’s also one thing that sets our approach apart from some of the other tools out there, where you never know exactly where your data ends up if you go to a consumer-facing AI chatbot.
Ross Katz: Yeah, that makes a lot of sense. As I understand it, you’re not training on the data—it’s a retrieval-augmented generation paradigm where you’re finding the most relevant documents, giving them to the LLM along with the query from the user, and using smart search and better question asking to get the best results possible for the user.
Alexander Junge: Exactly.
Ross Katz: I just want to tie a bow around FiiSH and the work that you did with Nordic BioVentures. How do they know when they’re done? Can you outline the process they go through and how they narrow down what they’re doing—you can do infinite amounts of research in this kind of environment. Your understanding of that would be interesting.
Alexander Junge: It very much depends on user to user, customer to customer. In this case they’ve been doing this before, and that’s our impression of a lot of the innovation teams in larger companies too. They have a fixed list or a timeframe—often it’s somewhat time-boxed—or a certain number of ideas they want to evaluate, and then we basically stick to that process and support it in partnership with our tools. Another aspect is how that’s going to look in the future. You could think about building these systems in the background and having them running all the time, just sending you a notification when something interesting comes up. I think we are a little bit off of that, given the technology and the stability. Language models these days are really cool and very powerful, but they’re not really at a stage yet where you can let them run independently for multiple hours and expect them to come back with a good answer. There are still some research breakthroughs to be done to make these kinds of things reliable. There is always going to be that touchpoint of regularly checking in with the expert, with the user to say, ‘am I going in the right direction, is this useful?’ then a human expert coming in and steering or opening the scope in terms of next steps.
Ross Katz: Other than user feedback—people from Nordic BioVentures coming to you and saying, ‘this was great’—how do you evaluate the quality of what you’re doing on an interaction-by-interaction basis so that you can improve the system over time?
Alexander Junge: That’s super important. That’s one of the key things to focus on—when I first got started in this, we still called it machine learning, now it’s all AI. You need evaluation datasets, otherwise you’re just flying blind. We have a mixture of public datasets. There are a few public datasets for the biomedical domain that we can use, and then it’s largely also internal datasets that we either annotated ourselves or synthesized, based on the kinds of user questions we’re seeing. That allows us to optimize the system in the right direction and make sure we’re tuning the right knobs.
Ross Katz: You’re not training on user data, but you’re certainly evaluating yourself on user data, I would assume.
Alexander Junge: Exactly. On user data or data that is inspired by user data. If you just use synthetically generated data that resembles user queries but isn’t exactly user queries, you can actually get very far. Once you have some usage—we’ve been in production with this system for more than a year now—we know exactly what kinds of questions people are asking and can use that to generate new questions that are similar to those without revealing a whole lot about our users themselves. The large language model providers, especially Anthropic in this case, have papers on how to do this in a privacy-preserving manner, and we take inspiration from that.
Ross Katz: That makes a lot of sense. Do you get concerned—since the evaluation datasets you’re using are so critical to how you tune the system—about over-indexing on the nature of the queries from your early users, the innovation teams, the business development teams, the smaller biotechs, the life sciences investors, and thereby not serving the needs of potential users? How do you think about that tradeoff?
Alexander Junge: Constantly. I think anyone operating in this space is paranoid around are we optimizing for the right things and are we measuring performance in the right way. That’s just the nature of the game. You need to be paranoid. Because the benefit—or the problem—of these AI models is that they will always give you some kind of answer. It’s up to you to make sure you have enough focus on the team and spend enough time actually looking at the data. Looking at regular user queries, seeing what activity is going on—not for labeling purposes, but to get an understanding of where the algorithm is failing, where it’s performing well, what are the problems in the system—that has a huge value. It’s about building a culture internally where we don’t leave it up to the system to operate by itself, but actually keep a close eye on observing the system as it’s running. That’s super important.
Ross Katz: Before we go there, one thing that everybody in this space has to notice is that you’ve got a lot of both startups and larger organizations promising to reimagine scientific research. You’ve got the big foundation model providers—OpenAI just announced, I think this week, the researcher they’re releasing. You’ve got Perplexity, NotebookLM, Elicit and these other AI-driven scientific research companies. You’ve got Future House doing agentic AI. We had Genomenon on, who help structure the scientific research in genomics and multi-omics. How do you think about differentiating Amass from the other organizations promising things in the same domain?
Alexander Junge: It’s a large number of types of organizations you just brought up. Starting with the Googles and the OpenAIs: we’re using a variety of AI models—language models, open-source models, embedding models, re-ranking models. We do benefit from the new advances and the models they put out. That makes our product better and that’s really cool. We’re building on top of the platform provided by the larger tech companies and the hyperscalers. You mentioned deep research—Perplexity is operating in that same space around consumer-facing search. These are really cool, we also use them in our day-to-day. But that’s where we see these kinds of tools falling short: they are consumer-facing products. What we are building is a professional tool made for life science professionals. That of course influences the kinds of search and information retrieval we are doing in the background. We can ignore travel websites—we don’t need to answer questions around where to go for holidays—and we can use that capacity to focus on the right datasets and index them in the best possible manner. Building a tool that is made for professionals that professionals understand. I think it’s really cool what OpenAI is doing, but every time I go back to ChatGPT every other month, the interface looks completely different. I looked at it recently—you can now actually choose from like 10 different language models to get an answer. Even as someone who would consider myself an expert in this space, it’s super confusing. What should I be using here? I just want to get an answer to my professional question. In the way we design our product, we’re not leaving that choice for the users to define which model to use. They don’t want to make that choice—they want to use the model that gets them the best answers. That’s how I would contrast ourselves to the big ones. Then there are solutions like Elicit—they’re doing really interesting work. For now they’re focused more on extracting smaller information into tables, focusing on systematic reviews, but across all the sciences. I think they’ve found a very interesting niche and can provide a lot of benefit for their users. We’re looking at them, reading their blog posts—I think we’re all learning from each other. Same for Future House. What I like about them is they very much have that nonprofit angle right now in the sense that they publish a lot of their research and don’t really care about spending time building consumer-facing products—they care about building a research team and they’re very vocal about that, and it’s really interesting. We’re looking at their code, looking at the papers, trying to take the best out of all these ideas into our system. Maybe the last thing is that we’re just getting started. The space of products and experiences to build is so large, even with the capabilities of the AI models we have right now, let alone what’s coming over the next years. We could be busy building good products for people for years to come without running out of ideas.
Ross Katz: Yeah, and it strikes me from hearing you talk about it that because the space is so large, the critical aspect is who are you serving and what do they want from the experience. The larger the population and the more diverse their interests, that impacts the design decisions you have to make. But with this seemingly narrower focus in terms of the datasets, choosing what your target customer looks like is really important. That brings me to another question—given that you’re serving people who know something about life sciences but are more breadth-focused than depth-focused, as I understand it you’re like a depth-focused assistant for breadth-focused people. Am I characterizing that correctly?
Alexander Junge: Yes. We can do both actually. We can be very broad, but we can also go deep with the user on a certain topic.
Ross Katz: Yeah, it strikes me that in the way that you structure your responses, you need to make assumptions about what people know and don’t know. How do you think about designing that experience into the system—giving people the right information at the right level for the decision they need to make?
Alexander Junge: I don’t think we have the final answers here. For us, what’s been really important is that ever since we’ve had a product that people could use, we’ve worked with many different kinds of organizations—all the way from academia with local universities and research groups to life science investors and biotech companies large and small. That gave us a very good view of how people see these tools, how they want to use them, what they trust in the answers, what they don’t trust. The first problem we faced were the classical hallucinations. Luckily we’ve solved a lot of that by now. But that was the first thing we needed to solve: how do we actually build trust in these answers? The trust we’re building and how we’re showing it right now is that everything we’re showing in terms of scientific information, we’re showing you the reference in a very transparent manner—‘this is the document it comes from and this is an extract from the document that says what we’re claiming here.’ We have this design pattern built into all the different products. What is this going to look like in a few years? I don’t know. There was this infinite scroll that someone invented at Twitter at some point when people were seeing these huge feeds. It feels like we need some kind of UX modes of interacting with these interfaces that work a little bit better, make it easier to gauge what’s out there. We’re doing our best right now, but again, it’s an open problem.
Ross Katz: You mentioned removing hallucinations and ensuring that all of your responses are grounded in actual papers, patents, and internal documents. Can you talk about how you went about implementing that, and as I understand it, that requires additional computational power? How do you manage the latency of the user experience with the rigor that you’re applying to the research that you’re doing?
Alexander Junge: We do focus more on rigor than providing fast answers. That’s also one thing we can do because we’re a professional-first tool—it’s fine if you wait a minute for an answer as long as the answer is reliable and trustworthy. That’s way more important than getting something very quickly. That’s our primary focus. How do we do this? Take the simplest example of a user asking one query. What we do in the background is rewriting that question—if there are spelling mistakes, etc.—into a proper question, and then translating that question into keywords. We do both: search based on the keyword level, but also in a vector embedding space. We have a big vector database of embeddings of research articles, patents, and other documents we’ve indexed. Searching both keywords and embeddings—a lot of people have seen that if you do both you get better answers and a more complete overview of the documents that might be relevant. We’re searching across all these different data sources and then doing a re-ranking across all of them. You might come up after this step with a hundred relevant documents. You don’t want to show 100 different documents in the very first answer—that’s a lot of text to read. We identify the most promising ones, take them into a language model, and at that point it’s the classical retrieval-augmented generation approach: ‘given this context, please try to answer this question.’ When we generate the answer, we’re not only generating one answer but actually generating different variants of answers and then generating a consensus answer on top of that, and finally cross-checking that consensus answer against the original documents to make sure the references we put in there are accurate and actually reflect what the original sources are talking about. There’s a lot of compute going on—it’s not just putting a question into a language model and getting an answer back. We spend a lot of time optimizing the system, burning through hardware, fixing what doesn’t work. Right now we’ve come up with a system that is pretty reliable and gives us very good answers in the cases we are seeing. There’s still space to make it better, as with every technology.
Ross Katz: Yeah, that’s interesting. In the example you gave, potentially you have 100 of the most relevant documents and you choose the top few. Is there a loop that happens where the user is saying ‘give me more information,’ so that you’re continuing along the digestion path of all the information that might be relevant? How do you think about allowing people to digest the large quantity of content that might be relevant to their search?
Alexander Junge: How we do this right now is that we show the user an answer that is digestible and then encourage the user to ask follow-up questions. In these follow-up questions we make sure that we summarize the conversation enough before we give it to the language model, to make sure we focus subsequent answers on the right context in the conversation. In every conversation, if you just go on and talk about all kinds of topics, it becomes hard to follow, and we saw the same kinds of problems early on. We built some mechanisms around summarizing things in a reliable and good way into the system.
Ross Katz: Can you talk about how the solution is hosted and how companies’ internal documents enter the ecosystem in a secure way?
Alexander Junge: Our solution is hosted on Microsoft Azure, and that means we are able to build on a lot of—we’ve only come across one company so far that is not using Microsoft. They’re very prominent in the life sciences. Building on top of the same infrastructure that our clients are already using enabled us to build privacy and security into the system from the very beginning and make our customers feel safe—we’re not putting your data into a third-party platform where it wouldn’t live otherwise. We’re benefiting from a lot of trust that has been built over the years with Microsoft specifically as a partner, and they’re also a good partner for us. We follow best practices around securing and encrypting data at rest and ensuring that only the people who really need access to data actually have it on our end.
Ross Katz: Yeah, that makes sense. The data’s already in the company’s cloud environment, you’re getting secure access to their cloud environment so that you can pull it into your system and interact with it without extracting it. That makes sense. As I understand it, the front end is a chat-based interface. Can you talk about the pros and cons of that? Because this is the baseline way we have for interacting with LLMs, but I’m imagining there are some ways in which that’s enabling the researcher’s use case you’re talking about, but also some ways in which it’s probably standing in the way.
Alexander Junge: Yes. We have multiple interfaces. Chat was the one where we started. The benefit is that it’s very broad—you get a good idea of what people would like to do in your system, where it’s not performing as well, where you can onboard other data. It’s a very flexible tool, but it’s also a huge surface area. What we learned from that is building what we call the Scientific Deep Dive, where the user doesn’t put in just one question—we encourage the user to explain in a paragraph what they’re looking for, what experiments they’ve been doing, what specific variants of a gene they’re looking at, the different factors that led them into the platform. Then before we generate any answer we say, ‘out of these 10 different dimensions we could be looking at, which are you interested in?’ The user selects those. It’s not a chatbot anymore at that point—it’s a simple multi-select interface. We break that down, figure out what topics the user is interested in, flesh those out into subtopics, and translate those into queries in our chat system and come back with a deep dive. The user never really asked one question at that point. We help the user figure out what they want to ask, generate those questions in the right way, and generate a long-form deep dive that a user can read and then ask very specific questions to. It’s almost like you can highlight a section in the report and say, ‘can you explain this in more detail’ or ‘that’s interesting but I also know about this other thing, can you please contrast your answer to this?’ It becomes more of a document-first way of interacting with the technology where the AI moves a little bit to the side. That is the kind of experience we’re building a lot more of right now. One of the interesting cases we’re seeing and building out together with multiple customers is grant writing. As a company you need to apply for research grants to get soft funding—from the European Union or different Danish government agencies, things like that, they exist in every country. The grant applications often relate to what is known in the world out there, who has done research on this, what do we have in our internal documents, and then synthesizing this information into a grant application. From a user experience standpoint it’s primarily: ‘please look at these sources here and give me a first draft of this document,’ and then as a user I get the chisel, not the hammer—working on this document together with the AI to say, ‘can you help me with this section? I don’t think this is concrete enough, it’s not really believable in the eyes of a reviewer. Can you help me make this more concrete and get my ideas across in a better way?’
Ross Katz: You named two very interesting interfaces that I’d like to touch on individually. The first one, you were talking about the chat interface as the entryway to this aided search paradigm—where the chat is the beginning of the interface surfacing to you based on what you’re saying, ‘this is what we think you’re looking for, could you clarify what you’re actually looking for using the topics that are there.’ Am I understanding that correctly?
Alexander Junge: Yep.
Ross Katz: Yeah, that’s really interesting. On the output side, obviously grant applications can run into the hundreds of pages depending on the nature of the grant you’re applying for. Are you helping organizations to generate hundreds of pages in an organized format, or how does the Amass platform really get someone to that document from a chisel rather than a hammer and nails?
Alexander Junge: For now we have not seen hundreds of pages of applications. I’m sure we’ll get there at some point. But it’s more on the two, three, four pages at a time that the user needs help developing. A lot of it in these writing applications is that you as a researcher very much know what the company is doing, but maybe you’re not the best at explaining it to a third party. Someone might be reviewing an application who doesn’t know anything really about your field, and the risk is confusing people and making your idea unbelievable by being way into the details instead of meeting the audience at the right level. That’s also one of the experiences we’re building into these grant writing applications—‘help me rephrase this for a certain audience,’ for example. Yes, these documents can be hundreds of pages long, but if you refine paragraph by paragraph that in itself has a big impact. Eventually we might be able to generate a first draft of very long documents, but again it’s always working hand-in-hand as a partner with the experts instead of generating all this automatically and hitting the submit button to send it to grant agencies. People want to have control over how they’re actually presenting their research project—it’s a huge reputation risk if you don’t review these things.
Ross Katz: Yeah, that’s really interesting. There’s on the part of the customer the desire to have an easy button to just do the thing all the way through the first time, but there’s also a desire to have control. Generating section by section in partnership with an AI that understands all of the literature but can also cast the results in the right format for the right audience is the right tradeoff to make there. I did have one more architectural straw man to throw at you. I’m imagining there are a lot of companies out there that have homegrown a solution that’s not that far away from what Amass is doing—they’ve grabbed a bunch of documents from various sources, chunked them up or tossed them all into ElasticSearch, run a search, and injected the results into an LLM so they can ask questions about it. What would you say to an organization like that about why Amass is a better solution than just dumping your data into search?
Alexander Junge: Having built these systems for a number of years now, it’s just not that easy. It’s very easy to get to the Twitter or X demo of your prototype, but then there’s a long tail of problems and failure modes that you’d never thought about when building the prototype that keeps you from creating actual impact and actual value with these tools. How we approach it is that if there’s an internal tool of some kind, we’re very happy to do benchmarks against those tools. Common questions, can we run these on different systems, can we compare. That makes the most sense and it’s way more informative than looking at public benchmarks that only look at an angle that’s maybe not super relevant for the given organization. Bring it on, I guess—let’s compare the systems. It’s an interesting time right now. We see a lot of homegrown solutions but we also see that a lot of them never really get into production in the sense that they’re never hitting any actual end users. That’s just what it takes to make these systems reliable. Maybe the last point here is that if you are working in a large organization and trying to do this internally, everything you’re optimizing for is the state of science that your organization has already looked at in the past. The expertise, the kinds of research projects running in your company at that stage. But a lot of these systems are built for new projects—forward-looking, venturing into new spaces—and that’s very, very hard to build with only internal data. It’s not really what you’re after. It’s the next thing that you don’t really know about.
Ross Katz: Yeah, that’s really interesting. We just talked about build versus buy for potential customers. Can you give us a little bit of insight into what you chose to build versus buy and why?
Alexander Junge: We are trying to only build what we really need to build and buy everything else. We’re running on top of databases—we’re using Postgres whenever we can, a traditional open-source database choice, same for our monitoring tools, language models, etc. We only really focus on building the stuff we truly need to build. We use existing AI models where they perform well enough for our domain. We’ve identified areas where they’re not good enough, and that’s where we’re now starting to build our own models specifically for the system. Embedding models—we’re starting to train our own embedding models because we’ve seen that as one of the primary failure modes. The existing embedding models, which have seen all kinds of domains from finance to economics, don’t really have a very deep understanding of the life sciences. We think we’re making a big impact by training our first models together with some good partners here in Denmark. I can’t really talk too much about our final results yet, but I’ll be able to do that soon.
Ross Katz: As I understand it, the rationale there is that you’re noticing in your vector search the results coming back are not as relevant as they could be, and the hypothesis is that’s because these models have been trained on a much broader set of data that does not have the context in the life sciences that would lead to the quality of results you would want. Am I thinking about that right?
Alexander Junge: Yeah. If you just think about a bench—that means a lot of things, but it has a very specific meaning in the life sciences. These kinds of things are what you can get out of training specific or fine-tuning existing embedding models towards a specific domain.
Ross Katz: Yeah, awesome. You touched on monitoring there and you made a point earlier I want to circle back to—how do you think about monitoring the systems so that you have all of the visibility you need to know when the system is operating as it should and when it’s not?
Alexander Junge: What we’ve built is—there are a lot of AI agent frameworks out there, the LangChains and CrewAIs and a lot of multi-agent frameworks that allow you to get to that prototype very fast, but you’re giving up a lot of control around what is happening behind the scenes. We’re not using any of those tools. We’ve built our own prompt flow, our own logic within the system in terms of how we use the different models, from the embedding model to the re-ranking model to the language model. That allows us to monitor and observe the exact right points and data points we want to look at and the exact right traces. We get that in a format we can actually understand—‘okay, it went here and here and here’—and we don’t end up in a crazy mess where there are five different AI agents supposedly doing something but it’s really hard to debug. We on purpose stay away from those systems that, in my experience, just become a big mess. We have a system that we understand, that we’ve built from the ground up, and we can monitor it very easily and keep track of additional things that should be relevant.
Ross Katz: The analog here sounds like a microservice architecture where you have different interactions and you’re monitoring each interaction point and have the traces through them.
Alexander Junge: Exactly. But we built it ourselves. We don’t rely on another tool or framework to take care of this—we say, ‘whenever we make a call to a language model we want to know we’re making this call, we want to know what the actual prompt is, we want to have influence over that,’ and then for logging we’re using third-party tools just to keep it in a database somewhere.
Ross Katz: Great. As we come to the end of our conversation, I want to ask you a few final forward-looking questions. The field of AI is moving very quickly. How do you think about evolving the Amass platform as models advance, as the tooling ecosystem evolves?
Alexander Junge: With models as they advance—there’s a big trend right now in inference time scaling, or test time compute, whatever you want to call it, reasoning models. That’s an interesting trend we are observing. We are in a lucky position in that we see the current models failing to some extent on the kinds of problems we are facing. We’re not in a scenario where GPT-4o is already able to answer 90%, 99% of the questions. We are going to be able to benefit from larger and better language models down the road. I really like open source—it’s a tide that lifts all boats—and the DeepSeek discussions over the last weeks, say what you want, but at least it’s great that this is happening in the open to a large extent. That is going to increase the speed of innovation in terms of improving those models, which they do need—they get lost on some trivial things sometimes. We’re benefiting from that and we will build those additional language models into the platform as we go along and as we see them performing well in our datasets.
Ross Katz: The new models come out and you’re constantly experimenting with them. Do you have an experimental process?
Alexander Junge: Not everything. There’s a handful of models we keep an eye on, and every now and then we run the experiment if it’s worth our time. That’s how we approach it right now. There’s also a lot of interactions with our users. It’s not that we say, ‘okay, this is the tool, please go ahead and use it and leave us alone,’ but every single time we demo our solution, we get at least 10, 20 questions back—‘could you be doing this, could you help us with writing press releases, could you help us with X and Y and Z?’ That’s a great source of information for us. We need to filter it down and prioritize, but that’s very much where we can say, ‘given the technology, what are the new experiences we can build on top of what we already have, and revisit maybe some of the things that didn’t work half a year ago.’
Ross Katz: Are there any upcoming features coming to the platform that you’re really excited about?
Alexander Junge: The scientific deep dive. The kinds of feedback we’re getting there are really interesting. That’s already done—we’ve had it in production for about half a year. If anyone’s interested in trying that out, deep dive going deep on life sciences research questions in a very reliable and cited manner, reach out to us via our website. In addition to this, a lot of the writing use cases—the grant writing we’re starting, we’re also exploring larger biomedical writing applications right now, for example in the regulatory domain, medical writing, things like that. If anyone is interested in explaining their problem to us, working with us, seeing some of the ideas we have, please do reach out. Also via our website—just sign up and you will get in touch with the founding team directly and we’ll get back to you.
Ross Katz: Given the work that you’re doing, how do you see scientific research and the workflow changing over the next three to five years?
Alexander Junge: Right now, given the two examples of where we’ve seen actual companies being built very much in partnerships with our tools, that’s really encouraging to see this impact right away, and it makes me super happy. One of the things I keep coming back to mentally when thinking about what might be happening in 10 years from now or five years from now is Dario Amodei, the CEO of Anthropic—he wrote this essay, ‘Machines of Loving Grace,’ around fall of last year. His first point, and the most extensive point in the essay, is actually that AI has the potential to contract a century or even longer of biological innovation into five to 10 years. He says AI is going to make that possible largely because it’s going to help humans understand science much, much quicker. I hope we’re going to get there. There are so many problems that biology needs us to help solve—all the way from tackling the climate crisis to treating diseases we have no cures for whatsoever right now, which there are plenty of. I do hope this vision will come to reality in some way, and the stuff we talked about today was one of the first steps in that direction.
Ross Katz: For sure. For listeners who are intrigued by Amass, you mentioned your website—where should they go if they want to learn more?
Alexander Junge: The website is amass.tech. Reach out to us there to the founding team directly. For me specifically, I’m on LinkedIn—Alexander Junge—find me there, connect, write me. From there you can also see links to my other social media profiles and website.
Ross Katz: Well, Alex, it’s been a pleasure having you on. I really appreciate the time and look forward to connecting down the line.
Alexander Junge: Yeah, likewise. Thank you.
Jason: And that’s it for this episode of Data in Biotech. If you enjoyed the episode, please subscribe, rate, or leave a review in your podcast platform of choice. See you next time.
