The Next Leap in Enterprise AI: Hyperscience Winter 2025 Release

Introduction & Market Context

Xabi Ormazabal (VP, Product Marketing):
Welcome to our webinar today. We’ll be talking about the Winter 2025 release—also known as R42 for those familiar with our numbering. I’m really excited to walk through some of the major trends we’re seeing in the market, how our technology is evolving, and what we’re learning on this journey with all of you.

I’m Xabi Ormazabal, VP of Product Marketing here at Hyperscience, and I’m joined by Chip VonBurg, our Field Chief Technology Officer, and Sundip Patel, Director of Product Management.

There has been an explosion of growth in AI over the last few years. It feels like it’s accelerating every quarter. Large language models are receiving massive investment—just recently I saw that OpenAI closed a deal with AWS for $38B in compute. It’s mind-boggling.

At the same time, we’re seeing several important trends about making all this AI actually relevant to businesses. A few examples:

Cloudflare:
We did a joint blog post with Cloudflare, a cloud security company. They highlighted concerns about large language models scraping massive amounts of website content without consent. It raises important questions about data ownership and what models are being trained on.

Meta & Scale AI:
There was a lot of industry buzz when Meta announced a 49% investment in Scale AI, a major data labeling company. Immediately afterward, OpenAI and Google Gemini announced they’d stop using Scale AI for data labeling. Why? Because whoever controls the data supply chain can infer how models are being trained and potentially gain an advantage.

This underscores the importance of having a business-grounded approach to AI and protecting your enterprise data.

Public sector policy changes:
HR 1 introduced new requirements for SNAP (food stamps) where all beneficiaries must now be recertified twice a year. That has created a massive surge in paperwork volume. Automation, AI, and ML-native applications are more critical than ever—what we call a “big beautiful bottleneck.”

Consumer lending:
With interest rates cut, we’re seeing a boom in loan applications. Customers are facing sudden avalanches of paperwork—another case where smart automation and AI are essential.

However, organizations are struggling to apply AI successfully. A recent MIT/Stanford study showed:

 
• 60% of companies evaluated enterprise-grade AI tools

 

• 95% of pilots were considered failures

 

• Only ~5% made it into production

The promise is huge, the need is huge, but applying AI in meaningful ways is still extremely difficult.

At Hyperscience, we believe documents sit at the heart of business processes. That’s why our focus on intelligent document processing matters so much.

And before we go further, I want to thank all of our customers. The recognition we’ve received recently is only possible because of you.

We’re proud to be the undisputed Leader in the inaugural 2025 Gartner Magic Quadrant for Intelligent Document Processing. This came directly from customers giving testimony to Gartner on the value they’re receiving. You can download the full report on our website.

Beyond Gartner, we’ve been recognized by Forrester, IDC Marketscape, GigaOm, and others. Again, that’s thanks to you.

Customers frequently ask: Should we build our own solution on a hyperscaler or use a best-of-breed platform like Hypercell?

When we compared fully loaded costs—staffing, infrastructure, features, accuracy—we found that customers achieve:

 
• 272% ROI over 5 years at 1M pages/year

 

• $1.9M NPV savings using Hypercell vs. building on a hyperscaler

We’re excited to continue building value with you.

With that, I’ll hand it over to Chip to walk through the platform and the R42 release.

Platform Overview & Differentiators

Chip VonBurg (Field CTO):
Thanks, Xabi.

When we look at Hyperscience, there are a few core components that make us stand out.

1. Our Models

We have 30+ proprietary, prebuilt models out of the box—for everything from preprocessing to transcription. These are foundational to the accuracy and performance customers expect.

2. Trainable Models & ORCA

Beyond the prebuilt models, we offer trainable models for machine print and handwriting. And then there’s ORCA, our newest and most advanced model—our multimodal reasoning and cognition engine.

We’ll talk about ORCA a lot today.

3. Human-in-the-Loop

We’ve built unique, deeply integrated human-in-the-loop capabilities that ensure accuracy while minimizing review overhead.

4. Blocks & Flows

Our block-and-flow architecture is one of the biggest strengths of the platform. It allows:

 
• Validation

 

• Enrichment

 

• External integrations

 

• Complex workflows

This flexibility is a major differentiator.

5. Admin & Reporting

Customers get full visibility into operational and business impact through built-in reporting.

Market Challenges Driving Our Themes

We organize R42 around three themes:

 
1. Understanding

 

2. Speed

 

3. Modularity

Here’s what we’re seeing in the market:

Accuracy challenges

Traditional IDP tools struggle to deliver what they promise. This pushes customers to explore alternatives.

LLMs as a “false easy button”

LLMs seem like an easy answer, but they generally fail to handle core IDP needs with accuracy, control, or consistency.

Compliance & redaction pain

Redaction workflows are often stitched-together, brittle, and not connected to core IDP pipelines.

Speed challenges

Many ML models are black boxes. Organizations often have no visibility into rework cost.

Modularity & cloud flexibility gaps

Legacy systems can’t benefit from modern multi-cloud architectures. Consistent model management is rare.

Release Cadence & Themes

We ship two major releases per year. R42 is our Winter 2025 release. R43 will be our Spring release. Between them, we provide patches and point releases.

The R42 themes are:

 
• Understanding: Major investments in ORCA

 

• Speed: Efficiency, lifecycle management, optimized human-in-the-loop

 

• Modularity: Expanded connectors and deeper ecosystem alignment

Before diving into ORCA, let’s look at how customers are using it today.

ORCA in the Wild: Customer Use Cases

Chip:

1. Insurance & Retirement Services — Long-tail documents

High variability, low-volume document types where training traditional models doesn’t make sense. ORCA’s zero-shot capabilities shine here.

2. Transportation Fintech — Long-tail, high variability

Similar pattern: hundreds of variations that appear infrequently. ORCA processes them faster and more cost-effectively.

3. Consumer Lending — Pay stubs at massive scale

This is extremely high volume. Using ORCA in a multi-model, AI-in-the-loop flow drives extremely high accuracy cost-effectively.

4. SNAP Eligibility — Public sector

With states now required to recertify beneficiaries twice a year, volume has doubled. Document variability is enormous. ORCA handles the complexity.

And with that, I’ll hand it back to Xabi to talk about SNAP.

Hypercell for SNAP

Xabi:
Thanks, Chip.

Hypercell for SNAP is a tailored solution designed for Supplemental Nutrition Assistance eligibility across all 50 states.

We help validate:

 
• Identity

 

• Income

 

• Residency

 

• Expenses

 

• Social Security numbers

 

• Citizenship

All to help states reduce payment error rates below 6%, avoiding penalties and protecting federal funding.

SNAP documents include a massive long tail—licenses, SSN cards, utility bills, passports, green cards, and more. Many are variable or unstructured.

This is exactly where ORCA’s zero-shot reasoning excels.

With that, I’ll hand it over to Sundip.

ORCA: Optical Reasoning and Cognition Agent

Sundip Patel (Director of Product Management):

We launched ORCA initially focused on extraction via layout-based prompting. Since then, we’ve expanded it dramatically.

Generalized Prompting

You can now prompt ORCA to do:

 
• Classification

 

• Summarization

 

• Reasoning

 

• Interpretation

 

• Flexible extraction

Deep Integration with Supervision

You can now see exactly where ORCA pulled information from on a document. This dramatically improves trust and reduces review time.

Document Chat with ORCA

Users can:

 
• Ask questions

 

• Compare content

 

• Summarize documents

 

• Perform generative search

All without data leaving your Hypercell environment.

DEMO 1: ORCA Zero-Shot Extraction & Reasoning

Sundip:

Driver’s license example

ORCA extracts:

 
• Name

 

• State

 

• Organ donor indicator

 

• Structured fields

All in zero-shot mode—no training required.

Complex table example

We fed a messy, complex document into a General Prompting block. We asked ORCA to extract:

 
• Countries of origin

 

• Product sizes

 

• Lot & bond numbers

 

• Quantity received

 

• Summary of what the document is

 

• JSON output format

ORCA correctly:

 
• Grouped similar values

 

• Parsed inconsistent formats

 

• Extracted from messy tables

 

• Produced the requested JSON

 

• Summarized the entire document

All zero-shot.

Combining ORCA with Specialized Models

Sundip:

Some of the strongest results come from combining specialized models with ORCA.

Examples:

 
1. Pay stubs (mortgage processing):
    
   • Consistency checks

    

   • ORCA as a second-pass validator

    

   • 90%+ end-to-end accuracy

 

2. Claims processing:
    
   • Field ID models locate fields

    

   • ORCA reads medical terminology

 

3. Freight forwarding:
    
   • ORCA handles handwritten notes, scribbles, and overwrites

Let’s look at the pay stub example more closely.

DEMO 2: Pay Stub Processing with ORCA in the Loop

Sundip:

We processed four pay stubs:

 
1. Standard ADP

 

2. County government

 

3. Mobile capture

 

4. Highly irregular format

The flow includes:

 
• Classification

 

• Identification

 

• Transcription

 

• Table extraction

 

• Data validation

 

• Conditional routing to ORCA

ORCA is only used where needed.

Documents 1–3 processed cleanly. The fourth was flagged for supervision due to ambiguous table values.

Composite Understanding Approach

Chip:

I want to emphasize how transformational this is.

We use:

 
• Preprocessing models

 

• Primary clusters (high-frequency formats)

 

• Secondary clusters (long-tail)

 

• Additional logic

 

• ORCA for final lift

This produces:

 
• High 60s–low 70s% STP before ORCA

 

• +10–12% accuracy gain from ORCA

 

• 90%+ end-to-end accuracy

Applied not just to pay stubs—this approach works for any document.

Redaction & Masking

Chip:

Three major use cases:

1. FOIA (public sector)

Automatically redact sensitive third-party data using:

 
• NLP

 

• Named entity recognition

 

• Signature detection

 

• Regex

2. Digital Twins / Synthetic Data

Replace PII with format-preserving synthetic data so you can use documents safely in lower environments.

3. BPO Outsourcing

Redact PII to safely offshore work and reduce operating costs.

Back to Sundip for a demo.

DEMO 3: Redaction & Masking

Sundip:

We show a FOIA document where:

 
• Petar’s data must remain visible

 

• All other individuals’ PII must be redacted

Flow includes:

i 
• Full-page transcription

 

• Redaction

 

• Supervision

 

• Masking or black-box redaction

In supervision:

 
• All candidate fields are listed

 

• You can keep/remove redactions with one click

 

• You can navigate pages via field list instead of scrolling

Output: a fully redacted, FOIA-compliant document.

Theme 2: Speed

Sundip:

1. Training Data Quality

We now surface inconsistencies between human-labeled ground truth and model predictions—making cleanup dramatically faster.

2. Faster Model Training for Large Layout Sets

Significant improvements for customers with high-volume, high-variance layouts.

3. QA & Peer Projection

Optimized long-form transcription UI + staffing projections.

Theme 3: Modularity

Sundip:

New connectors:

 
• Google Cloud Storage

 

• Azure Blob Storage

In addition to AWS.

Document Chat with ORCA

Ask questions, summarize, compare content—all inside Hypercell with no data leaving the environment.

Back to Xabi.

Recap & Q&A

Xabi:
Thanks, Sundip. Let’s move into Q&A.

Q&A Highlights

(All corrected speakers)

Q: How does ORCA compare to other GenAI tools?
Chip: ORCA outperforms many general-purpose GenAI models because it’s tuned specifically for IDP. We’d love to show you a head-to-head comparison.

Q: How difficult is it to set up a digital twin environment?
Chip: Easy—especially if you already have Hyperscience models deployed. You reuse the same field detection to redact or replace sensitive data.

Q: Do you track accuracy, bias, precision, etc.?
Sundip: Yes. You can dial in automation thresholds, review metrics, and monitor performance across your entire pipeline.

Chip: And ORCA is a first-class citizen—same thresholds, confidence scoring, and oversight as all other Hyperscience models.

Q: Can ORCA extract full tables?
Sundip: Yes. Accuracy varies by document, but the generalized prompting example shows how powerful it can be. We’re happy to evaluate your documents.

Q: Does Hyperscience have system downtime alerts?
Xabi: Yes. Public status page is at status.hyperscience.net. We also have internal tools and solutions architects who proactively notify impacted customers.

Q: Does ORCA return bounding box coordinates?
Chip: Yes—generalized areas today, improving over time. We’re already ahead of many others in this space.

Q: How is ORCA priced?
Xabi: Pricing is based on:

 
1. ML components you select (ORCA, trained models, synthetic, etc.)

 

2. Blocks used (standard vs. advanced)

 

3. Volume (pages/year)

 

4. Infrastructure (on-prem, Hyperscience-hosted, FedRAMP)

ORCA uses GPUs, so right-sizing compute is part of scoping.

Closing & Calls to Action

Xabi:

To wrap up, here are three key next steps:

1. Explore the R42 / Winter 2025 Release Page

Includes all new features, demos, and documentation.

2. Download the Build vs. Buy White Paper

Deep dive into the 272% ROI analysis.

3. Talk to an Expert

Book time with our team.

In ON24, you’ll also see related content in the paperclip widget.

Bonus: Try our new ORCA gamified experience—see if you can extract invoice data faster than ORCA.

Thank you so much for joining us. You’ll receive an email with all links and resources after the webinar.