Spring ’26 Release: From IDP to Intelligent Inference

Sarah Westhoven: Hello, and welcome to the Hyperscience Spring 2026 release webinar. We really appreciate you taking time out of your day today to spend some time with us and hear what is new with Hyperscience. We have a lot to go through today, so we will kick it off.

To introduce our speakers for today, first up, we have Boyan Kelchev, our VP of Product. Boyan joined Hyperscience in 2019 and leads the product management function, overseeing product management, product design, and technical program management functions.

Next up, we have Xabi Ormazabal. Xabi leads the product marketing function at Hyperscience and brings significant expertise from leading international teams at Google Cloud, Salesforce, Dropbox, and other B2B technology companies.

Finally, we have Chip VonBurg, our Field CTO. Chip has over 25 years of experience designing, building, and leading teams and technology programs. He works closely with Hyperscience customers to make sure they achieve their business objectives and deliver measurable business results.

And I am Sarah Westhoven, Lead Product Marketing Manager at Hyperscience. Thanks for joining us today, guys.

To review our agenda today, first Xabi is going to take us through key trends shaping the IDP market and how HyperCell is helping customers balance accuracy, automation, and affordability through inference layering. Chip will fill us in on a current client success story that illustrates how inference layering works. Then Boyan will take us through the ORCA spring release updates and walk through a demo with Chip talking about some of the exciting new ORCA use cases and client innovations we are seeing. Finally, Boyan will wrap up with some exciting roadmap and release cadence announcements.

We will leave some time at the end for questions. You can drop them in the Q&A chat and we will answer them either in the chat or I will read them out loud to our speakers during the Q&A portion.

Let’s talk about our three themes for today. First, differentiation. HyperCell gives you a production-grade orchestration layer that is able to intelligently route your document processing workloads to the right model for the right task. High-volume, straightforward transactions go to efficient CPU-based models, while complex, high-stakes documents shift to our ORCA VLM for deep reasoning. We are going to show you today how that results in a balance between accuracy, automation, and affordability, and what that inference layering looks like in practice.

Second, innovation. We talked about showcasing our ORCA VLM framework, release items, and walking you through a demo, talking about how we have expanded hardware and frontier model support significantly, and how HyperCell brings those models into your workflows with the built-in thresholding, fine-tuning, governance, and observability that we are known for.

Finally, acceleration. We know predictability is extremely important to enterprise environments. Alongside innovation, we will talk about how we are evolving our release schedule and our roadmap to give you the continuous innovation you need to stay competitive with the stability that your production environments demand.

With that, we will start off the first section with differentiation. Xabi, I will hand it over to you.

Xabi Ormazabal: Thanks so much, Sarah. We are at a very exciting time in our industry right now with a lot going on. Obviously, as an ML-native AI platform for intelligent document processing, we are continually focusing on and analyzing what is going on in the market. Here are some of our observations in the last few months of what is shaping up to be some interesting trends.

First, we have been going to some major trade shows, working with our partners to better understand the evolution of their offerings and how we fit into this whole ecosystem of AI and the different solutions that organizations are delivering. We were recently at NVIDIA GTC. Jensen Huang had an amazing keynote with really interesting trends.

We picked up on a few that are very relevant to what we do, starting with the inference inflection point. Jensen talked about how AI is moving beyond simple generative tasks to really being able to think, reason, and act in a much more effective manner. The bar is being continually raised for what organizations are expecting out of agentic experiences and for AI to resolve effectively. This need for better and more effective inference is driving a huge need for AI compute.

That flows into the next point: organizations are going to start thinking about token consumption as the new commodity. Different departments will have budgets for how many tokens they can use within a year, whether you are in engineering, finance, or go-to-market. There will be a continuous need for organizations to differentiate to be able to consume that AI at scale efficiently. It is going to help you differentiate competitively and get the resources you need to complete your projects effectively.

We will talk later in this presentation about some of the improvements we have made to the different GPUs and models we are supporting. One thing that our teams are obsessed with is continually refining and improving the throughput with which the GPU models can actually process more and more documents with our ORCA model. We are very acutely aware of the need to make that AI spend be at scale but also be efficient.

Finally, at Google Cloud Next, a phrase came out from Thomas Kurian’s keynote that was really interesting, where he talked about how AI agents need a lot of context clarity to be effective, and basically, you need to avoid context pollution. Context pollution means you need accurate, ground-truth data focused on the task that an agent needs to execute. That is critical, and that is something we bring a lot of value to in the document processing arena specifically.

We have published a few blogs recently talking about the need for AI to have that deep contextual understanding to be impactful. One example was around being able to redact and mask data at scale, such as for the Freedom of Information Act workloads we are seeing with the USCIS and other organizations.

We also mentioned how AI build costs are spiraling out of control. A lot of organizations are faced with the dilemma of whether to buy a purpose-built platform for what they need to achieve or build it themselves. What we are hearing consistently across many customers and prospects is that doing it yourself on a hyperscaler has a lot of unpredictability in terms of cost and complexity. We published a white paper, which you can find on our website, regarding how Hyperscience can be much more cost-effective and deliver a higher ROI by building on our platform rather than building from scratch with a hyperscaler.

AI needs to deliver demonstrable ROI. Our CTO for the public sector, Trevor Diallo, published a great blog post talking about how trust, transparency, and technology that delivers are really key imperatives for the public sector. We are seeing this need to drive demonstrable ROI to be very important, not just in the public sector, but in the private sector as well.

Another key shift is around the concept of human-on-the-loop. As a machine learning-first organization, we always talk about human-in-the-loop as a key mechanism for quality assurance, understanding the accuracy of your models, and refining their performance. But now we have evolved to this concept of human-on-the-loop, where we want to give greater oversight to all of these processes, models, and agents working as a whole in order to drive better outcomes. We see that as a big requirement and something that is important for us in this concept of inference layering.

One of the ways we approach complex problems with intelligent document processing is a series of techniques we call inference layering. We want to drive home the point that with a bigger toolkit of solutions, you can apply the right model to the right task at the right time. It guarantees you faster extraction, better reasoning, and better governance in your projects.

The three things that underpin this for us are accuracy, automation, and affordability. We want to make sure that you are able to leverage not just probabilistic approaches with generative AI, but also deterministic approaches where required. We want to make sure that you can actually get more done with less human interaction. If you initially needed human-in-the-loop to refine your models, how do we minimize that and bring other tactics such as AI-in-the-loop? We also want to make sure this is affordable, predictable, and scalable for your organization.

Before I go into a deeper explanation of inference layering, Chip, do you want to touch on what you are seeing in terms of customer interactions across accuracy, automation, and affordability?

Chip VonBurg: This is almost like strapping a turbocharger on the solution that you have in place today. It allows you to take accuracy to the next level in a cost-optimized fashion. It is allowing customers to see accuracy and automation rates with projects that were unachievable in the past, and it does it in such a way that we are not using a sledgehammer for a finishing nail. It is changing the way we are thinking about projects, how we go after them, and the tools we use to execute them. We are going to talk about a couple of those use cases as we get further along, but it is a big game-changer for sure.

Xabi Ormazabal: When we talk about this concept of inference layering, this visual might help, showcasing a layer cake starting from the bottom and going up to the top. If the whole shape on the screen is the entire corpus of all your documents, we want to start and progressively apply different techniques to understanding and extracting the information, covering core IDP tasks like identification, classification, and extraction.

At the first level, we handle pre-processing and classification. This includes all the different techniques we have within our pre-built models that allow you to de-skew, de-noise, rotate, and correct visual anomalies on documents, understand what those documents are, and apply external models like Gemini Flash. That first layer of tactics will handle about 10% of the volume of the documents you are processing.

A key part of this is clustering the documents, understanding their makeup, and grouping them by difficulty or complexity. If you have a primary set of clusters, you go through processes for identifying fields, creating those fields and templates, and applying models such as Beaker or Optical Intelligent Character Recognition. These specialized models allow you to execute tasks around extraction and pulling out information, which we can supplement with other models like Gemini Flash. That takes care of a much broader percentage of the document volume.

Then you continue to work on the secondary clusters and apply more advanced techniques using different code blocks. We will talk about blocks and flows in our architecture and the ability to concatenate the logic you need. As you move on to more sophisticated reasoning and detecting anomalies, each of these layers addresses a progressively more difficult level of inference, handling a specific percentage until you complete everything.

At the very top, we have the long tail, where we are able to apply human-in-the-loop and AI-in-the-loop, meaning we can recursively call these models to help us refine the final tranche or most difficult level of documents. Chip, I know you have seen this applied rigorously with our customers. Is there anything else you want to land around this concept of inference layering?

Chip VonBurg: First off, I think this is a great graphic. When you talk through this verbally, people get it, but the visual lays out exactly what is out there. The ability to intelligently use different models to go after different parts of the document population is a game-changer. It allows you to add layer upon layer of intelligence to take accuracy and automation through the roof. Let’s take it to the next slide and I can talk about one of these use cases in particular.

Xabi Ormazabal: The only other thing I would add is that at each level, we also do complex checking of values on the document itself. If we are talking about pay stubs, it can look at what the payments and retentions are in the current pay period versus year-to-date to see if those match up. That complex level of in-document calculation and reasoning drives the understanding that we successfully extracted the data or alerts us to move to the next layer if we didn’t. Over to you, Chip.

Chip VonBurg: We are seeing this more and more. We have a number of these out there now where we are using inference layering, and it is starting to become one of our standard deployment methods.

A recent implementation was with one of the Big Four consultancies, specifically going after invoice and tax automation. This particular firm was trying to read invoices from all of their customers to extract tax information and categorize what those invoices were about. For any of the Big Four, getting the data into their system is the first step toward applying their logic and delivering services to their customers. Being able to accurately and quickly get that data is a big value to them.

The challenge is that there is a gazillion different invoice formats out there. Because you are looking at this from the angle of a consultancy, the typical groupings by industry or vendor go out the window. The challenge is much greater than what you might see for a typical IDP case at a single customer; it is heavily amplified. They needed to read this data quickly and accurately across unlimited variations of documents.

They tried legacy OCR tools and even tools from the hyperscalers. The hyperscalers did better than legacy OCR, but the question was: at what cost? That is when Hyperscience came into the picture.

First, notice the field-level accuracy settings we call out on the right-hand side of the chart. This feature allows you to give more importance to different fields on the document. In this use case, fields like tax and total are more important than fields like terms. By tuning where the models are looking and where they put the performance, we are able to drive the required accuracy levels for the high-priority fields while keeping the automation level very high. We get as much data as accurately as possible with as few touches as possible.

The second thing we did is exactly what Xabi was talking about: using agent-in-the-loop. In this use case, we are using ORCA in the loop—our VLM—alongside the hyperscaler’s inference tool. We bring all of those into the same single orchestration layer within our flow, allowing us to call different models based on specific needs. This adds agentic layers to the process. Instead of taking the traditional approach of taking one shot at it and then throwing a human at it, we take several shots. We try the first model, and depending on the fields and documents, we will try the second and potentially a third. This brings accuracy way up and keeps automation levels very high.

At the end of the day, the winning combination was using the features we know and love in Hyperscience—like field-level accuracy settings and the orchestration platform—and bringing all of these inference layers in to achieve automation while keeping costs reasonable.

The size of some of these firms is giant, and the amount of data they deal with is massive. Xabi brought up the cost of tokens, and you are going to hear it time and time again. Tokens are one of the new currencies out there. This firm is approaching over a trillion tokens per month. That is trillion with a “T.” Helping them optimize that token spend is a big deal.

Because this is all built into the same orchestration layer and delivered as a single project unified by the flow behind the scenes, we were able to hit a go-live date in under three months. Not only did we get them accuracy and cost savings, we did it quickly.

Lastly, using both the compute and inference layers helps them use the cloud commits they already had with their hyperscalers. They can leverage the spend they already have there to get the results they need. We will touch on a couple of other use cases, but hopefully, that gives you an idea of what we are seeing out in the field. With that, Sarah, we will transition things over to Boyan.

Boyan Kelchev: Hi, everyone. In this first theme for innovation, we will start talking about the features we are bringing to market, and I will walk you through a couple of demos.

Many of you are already using our most powerful solution: our Optical Reasoning and Cognition Agent, ORCA. But since some of you may not have tested it yet, we will start with a couple of definitions. We released ORCA a little over a year ago, in early 2025. Over the past several releases, we have added a lot of capabilities to the model and the underlying framework.

The best way to think about ORCA is in the context of the overall framework. There is a powerful foundation model behind the scenes that has reasoning capabilities and functions out of the box for many use cases. Natively integrated with that model is the harness around it, which allows us to provide the safety and confidence in generative models required for critical enterprise environments. This includes features like dynamic confidence thresholding, quality assurance, and human-in-the-loop. Many of you are already using ORCA for out-of-the-box classification, zero-shot extraction, and AI-in-the-loop techniques, meaning a subset of previous human-in-the-loop tasks can now be fully automated. Because it is a larger model, it can tackle unique extraction tasks that are not easy for non-generative models, aided by generalized prompting, decisioning, and reasoning capabilities.

In this latest release, we are delivering a long list of capabilities, but there are three we want to highlight today centered around the ease of use of ORCA. Now that the value of the model and framework has been proven, we have made it incredibly easy to start by yourselves without help from Hyperscience. You can deploy the model with a one-click ORCA install, use our out-of-the-box ORCA subflow to start testing new documents immediately, and fine-tune the model within the platform using a business user-friendly interface tied to our model lifecycle management capabilities. I will do a short demo to highlight what those look like.

(Boyan shares his screen)

Let’s start with the ORCA one-click install. Those who have used ORCA before might have worked with Hyperscience representatives to set up the large model, especially if on-premise. In the latest release, you go to the administration page and select the Assets tab. You will see a tile for ORCA, and clicking the install button brings up a menu with instructions on how to deploy it. There are two options for provisioning the model. For anyone on our SaaS offering, this is all transparent and set up automatically. For on-premise or private cloud environments, these two options allow any private cloud or air-gapped environment to set up ORCA with ease.

The first option, the artifacts repository, is for completely air-gapped environments that cannot open an internet connection even briefly. In that case, you provide a local network connection link to the model location so the system can install it. Most of our private cloud clients will use the second option: fetching through Cloudsmith. All you have to do is provide a Cloudsmith API key provided by Hyperscience. When you click start, it begins the process of installing the model. It takes about 15 minutes, and you are off to the races.

Let’s jump to the second walkthrough: the ORCA out-of-the-box subflow. Composability is an important design characteristic of our platform. We have architected our flows so you can take entire flows or parts of them and create reusable components across departments or working groups. We provide a subflow that uses ORCA out of the box so you can start testing new use cases immediately.

On the flows page, if we search for ORCA, you will see the “document processing with ORCA subflow.” It looks very similar to the standard document processing flow that comes with the platform. You have the same capabilities for pre-processing, file filtering, machine collation, machine classification, and natively integrated human-in-the-loop tasks. The key difference is the optical reasoning and cognition agent block, which allows ORCA to be used instead of having to train FieldID or other models. This is backed by our flexible extraction interface and an optional quality assurance flow to understand the exact accuracy of the model.

To enable it for new use cases, we recommend duplicating the top-level document processing flow, which is done with a single click. We will click into our webinar flow and show you how to swap out the standard subflow for the new ORCA out-of-the-box subflow. As you click into the block, there is an option under “flow identifier” on the right-hand side menu. All you have to do is select “document processing with ORCA subflow” for the flow to use that capability. You can also toggle the quality assurance flow option. That is all it takes to run new document types through the flow and see the out-of-the-box results of the base model.

In cases where you have high business accuracy SLAs or unique extraction requirements that make a use case difficult, you can fine-tune the base model. Clicking into the Models page under VLM field extraction, you click the “Create model definition” button. Think of a model definition as a version of the underlying vision language model fine-tuned for a particular document type. You click Create model definition and select a layout—in this case, a medical history form. Under training data, you can upload documents to fine-tune ORCA. We have seen great results with as few as 40 documents. There are many options for filtering through the training data to adjust annotations. Training and fine-tuning the model is a single click.

The annotation experience is very similar to our Field ID and Table ID experiences, so no retraining is needed for your business users. You can see the fields for the layout on the right-hand side and adjust existing annotations easily. If a field is not pre-annotated, it is a single click or a drag-and-drop to annotate it. We have added capabilities to the model lifecycle management experience, such as adding tags to different documents within the training set to make it easier to add notes and filter.

In the Overview tab, you see performance expectations for the model, showing projected automation at different target accuracy levels for both the live model and the model currently being trained. On the projected automation curve, you can see what automation to expect at different target accuracy levels, such as a 95% target. Our reporting capabilities are completely integrated with this to track performance during processing. Lastly, the History tab allows you to see the different versions of the model you have fine-tuned, giving you the ability to roll back to a previous version with a single click.

To summarize the demo, we saw the ORCA one-click install, which deploys the model in minutes on-premise. We saw the ORCA out-of-the-box subflow for immediate testing of new use cases. Finally, we saw how you can fine-tune the model to meet high business accuracy SLAs using standard annotation and training experiences, but with far fewer document requirements to get better results. With that, let’s take it back to the rest of the content. Over to you, Chip.

Chip VonBurg: Thanks, Boyan. These are some specific use cases we are seeing across different industries that directly tie back to the points Boyan made during the demo.

The first use case is a customer in retirement services who needed to process a long tail of supporting documents like checks, driver’s licenses, passports, and transfers. The challenge was handling that long tail, and we leveraged the zero-shot capabilities of ORCA. Utilizing ORCA allowed them to process those documents without the upfront time investment required to build out individual models, resulting in faster time-to-value.

The second use case is in life insurance, where there was a long tail of over 3,500 layouts required to meet implementation needs. Once again, we used the zero-shot capabilities of ORCA to process those documents without upfront model-building investments. The common theme across these two use cases is zero-shot capabilities and the savings from going live with a larger group of documents immediately. In the first use case, this yielded nearly $850,000 in savings, and in the second, over 1,100 hours of time saved.

The last use case is in mortgage processing for pay stubs, though we are doing this with a few other documents in mortgage processing as well. In this case, we used ORCA in-the-loop to add that agent-in-the-loop experience, driving accuracy rates far above what they have seen in the past. Those incremental accuracy gains mean real savings for the customer, potentially equating to millions of dollars of increased revenue because they can approve and fund loans much faster, directly impacting their bottom line.

The last point I want to make is that inference layering is not just about being model-agnostic. We have talked about ORCA and Gemini, but we are continuing to extend the models and the GPUs we support. This gives our customers the flexibility to optimize the overall total cost of ownership and drive accuracy rates at the best optimized cost. We are continuing to expand our GPU support alongside model support. Boyan, back to you.

Boyan Kelchev: Before we jump to the roadmap and highlight what we are working on in the next nine to 12 months, I wanted to highlight three different use cases where we have worked with a number of you on novel approaches: inference layering, AI-in-the-loop techniques, and combining our orchestration layer, model harness, and ORCA. We have design partners for these, but we are looking for additional partners interested in testing these capabilities.

The first is out-of-the-box capabilities for tabular extraction, leveraging new solutions on the platform. The second focuses on our orchestration layer and how we augment third-party models to make them safer and closer to deterministic by utilizing our built-in guardrails. Last but not least, we are addressing dynamically generated and highly variable forms. With new tooling within the platform, our machine learning team is experimenting with new approaches and seeing great results.

Let’s move to the roadmap. I will be brief, but we wanted to highlight our focus for the next nine to 12 months. We continue to structure our roadmap across three themes: advanced machine learning, software AI infrastructure, and end-to-end automation and processing.

Within advanced machine learning, we are continuing to invest in our specialized models, which provide a cost-effective, accurate, and deterministic solution. However, we are shifting a lot of investment to our newest solutions, focusing less on the models themselves and more on the harness and orchestration layer that allow the smart, intelligent combination of different tools within the toolbox. We are looking at out-of-the-box table extraction, enhancing our out-of-the-box classification harness, and hardening AI-in-the-loop techniques so that tasks previously requiring supervision and quality assurance can be performed by agents within the platform.

Within the second theme, the focus is on making sure the largest models perform effectively at a very high scale. This includes vertical auto-scaling for high-memory workloads and low-latency processing optimizations. We are also investing in the components used to build these solutions, making third-party models first-class citizens within the platform for top-tier management.

Within the third theme, we are using agentic coding extensively internally within Hyperscience, and we expect to make that available within the platform in the near future for easier flow development. We are also adding capabilities to our reporting layer, looking toward an insights engine in the future.

A very quick update on our new release timeline. For most of you, this is a seamless change. Starting with this release, we are changing our release schedules to align across all deployment platforms, including on-premise environments and SaaS. We are making the latest minor release in each major release line the cross-platform release. This release, 42.3, is now available for all clients, including on-premise customers. In the fall, the latest minor release of release 43 will also be available to all. This accommodates release schedules and blackout dates while introducing additional stability within the on-premise releases. Bi-weekly service updates and bi-monthly SaaS innovation releases remain unchanged. We are looking forward to this new cadence.

Sarah Westhoven: Great. Thank you, Boyan, and thanks to Xabi and Chip as well. Here we have linked a few resources, which are also available in the resources section in the webinar portal.

We have a lot of Hyperscience customers on the webinar today. If you found value in working with Hyperscience, we would love it if you took a couple of minutes to leave us a review on Gartner Peer Insights. We really value honest feedback, and it means a lot to our team. Feel free to scan the QR code here, and as a thank you, we will send you a gift card.

Let’s move on to Q&A. We had a couple come through in our chat, and we have time for one or two. The first question: “Will standard use cases like pay stubs and invoices work with the ORCA base model?” Chip or Boyan, that is probably for you.

Boyan Kelchev: Yes, those use cases work with the base model. The base model works for any document type, but the devil is in the details. Depending on the number of entities you are trying to extract, as well as the complexity and variation within your document estate, there might be a need for either fine-tuning ORCA or layering inference with the techniques Chip and Xabi mentioned. The answer is yes, but providing the additional capabilities to optimize these use cases is a priority for us as clients continuously look to extract more data and introduce more variation.

Chip VonBurg: To add to that, these are use cases we see from customers all the time, so we have a lot of knowledge and expertise around them. We can help accelerate those use cases for you, whether we are using straight ORCA, ORCA in the loop, or another combination.

Sarah Westhoven: Excellent. Thank you. Next question: “Our IDP environment is on-premise because of regulatory requirements. Can we use ORCA on-prem?” Boyan, that is likely for you.

Boyan Kelchev: Yes, absolutely. ORCA can run completely in air-gapped environments. For the installation, we offer an option to use a short-lived internet connection to download the latest model as we continuously update it. However, that is not a requirement. You can operate in a completely air-gapped fashion by placing the model binary in a local network location where the platform can install it.

Sarah Westhoven: Excellent. Boyan, I think this next one is for you as well: “My understanding is that ORCA is a zero-shot model with no training required. What benefit is there to fine-tuning it?”

Boyan Kelchev: ORCA is a zero-shot model, so you can start testing any use case right away with the base model. However, in some cases, the base model might not give you sufficient results for your specific business accuracy SLA. In those cases, you have the option to use fine-tuning within the platform without any help from us to optimize the model for that particular document type. With generative models, it is commonly seen across the market that real enterprise use cases require fine-tuning in certain scenarios. We wanted to make sure that is available as an easy solution within the platform, rather than requiring you to deploy machine learning engineers or data scientists to do so.

Sarah Westhoven: Excellent. I know we are at time, so we will wrap it there. Thank you so much for taking time out of your day to hear what is new with Hyperscience. If you have any questions, please feel free to follow up with us after the webinar. Thank you very much.