Automating advanced doc processing: How Onity Group constructed an clever resolution utilizing Amazon Bedrock

Within the mortgage servicing business, environment friendly doc processing can imply the distinction between enterprise development and missed alternatives. This submit explores how Onity Group, a monetary providers firm specializing in mortgage servicing and origination, used Amazon Bedrock and different AWS providers to rework their doc processing capabilities.

Onity Group, based in 1988, is headquartered in West Palm Seaside, Florida. Via its major working subsidiary, PHH Mortgage Company, and Liberty Reverse Mortgage model, the corporate gives mortgage servicing and origination options to owners, enterprise purchasers, buyers, and others.

Onity processes hundreds of thousands of pages throughout a whole lot of doc sorts yearly, together with authorized paperwork akin to deeds of belief the place essential info is commonly contained inside dense textual content. The corporate additionally needed to handle inconsistent handwritten entries and the necessity to confirm notarization and authorized seals—duties that conventional optical character recognition (OCR) and AI and machine studying (AI/ML) options struggled to deal with successfully. Through the use of basis fashions (FMs) offered by Amazon Bedrock, Onity achieved a 50% discount in doc extraction prices whereas bettering general accuracy by 20% in comparison with their earlier OCR and AI/ML resolution.

Onity’s clever doc processing (IDP) resolution dynamically routes extraction duties primarily based on content material complexity, utilizing the strengths of each its customized AI fashions and generative AI capabilities offered by Amazon Net Providers (AWS) via Amazon Bedrock. This dual-model method enabled Onity to handle the dimensions and variety of its mortgage servicing paperwork extra effectively, driving vital enhancements in each value and accuracy.

“We would have liked an answer that might evolve as rapidly as our doc processing wants,” says Raghavendra (Raghu) Chinhalli, VP of Digital Transformation at Onity Group.

“By combining AWS AI/ML and generative AI providers, we achieved the right steadiness of value, efficiency, accuracy, and pace to market,” provides Priyatham Minnamareddy, Director of Digital Transformation & Clever Automation.

Why conventional OCR and ML fashions fall quick

Conventional doc processing offered a number of basic challenges that drove Onity’s seek for a extra refined resolution. The next are key examples:

• Verbose paperwork with knowledge components not clearly recognized
  • Difficulty – Key paperwork in mortgage servicing include verbose textual content with essential knowledge components embedded with out clear identifiers or construction
  • Instance – Figuring out the precise authorized description from a deed of belief, which is perhaps buried inside paragraphs of legalese
• Inconsistent handwritten textual content
  • Difficulty – Paperwork include handwritten components that adjust considerably in high quality, fashion, and legibility
  • Instance – Easy variations in writing codecs—akin to state names (GA and Georgia) or financial values (200K or 200,000)—create vital extraction challenges
• Notarization and authorized seal detection
  • Difficulty – Figuring out whether or not a doc is notarized, detecting authorized court docket stamps, verifying if a notary’s fee has expired, or extracting knowledge from authorized seals, which are available a number of shapes, requires a deeper understanding of visible and textual cues that conventional strategies would possibly miss
• Restricted contextual understanding
  • Difficulty – Conventional OCR fashions, though adept at digitizing textual content, usually lack the capability to interpret the semantic context inside a doc, hindering a real understanding of the data contained

These complexities in mortgage servicing paperwork—starting from verbose textual content to inconsistent handwriting and the necessity for specialised seal detection—proved to be vital limitations for conventional OCR and ML fashions. This drove Onity to hunt a extra refined resolution to handle these basic challenges.

Resolution overview

To handle these doc processing challenges, Onity constructed an clever resolution combining AWS AI/ML and generative AI providers.

Amazon Textract is a ML service that automates the extraction of textual content, knowledge, and insights from paperwork and pictures. Through the use of Amazon Textract, organizations can streamline doc processing workflows and unlock helpful knowledge to energy clever functions.

Amazon Bedrock is a totally managed service that gives a alternative of high-performing FMs from main AI firms. Via a single API, Amazon Bedrock gives entry to fashions from suppliers akin to AI21 Labs, Anthropic, Cohere, Meta, Mistral AI, Stability AI, and Amazon, together with a broad set of capabilities to construct safe, personal, and accountable generative AI functions.

Amazon Bedrock provides you the pliability to decide on the FM that most accurately fits your wants. For IDP, widespread options use textual content and imaginative and prescient fashions akin to Amazon Nova Professional or Anthropic’s Claude Sonnet. Past mannequin entry, Amazon Bedrock gives enterprise-grade safety with knowledge processing inside your Amazon digital personal cloud (VPC), built-in guardrails for accountable AI use, and complete knowledge safety capabilities which can be important for dealing with delicate monetary paperwork. You’ll be able to choose the mannequin that strikes the best steadiness of accuracy, efficiency, and price effectivity on your particular software.

The next determine reveals how the answer works.

1. Doc ingestion – Paperwork are uploaded to Amazon Easy Storage Service (Amazon S3). Importing triggers automated processing workflows.
2. Preprocessing – Earlier than evaluation, paperwork endure optimization via picture enhancement, noise discount, and structure evaluation. These preprocessing steps assist facilitate most accuracy for subsequent OCR processing.
3. Classification – Classification happens via a three-step clever workflow orchestrated by Onity’s doc classification software. The method outputs every web page’s doc sort and web page quantity in JSON format:
   1. The applying makes use of Amazon Textract to extract doc contents.
   2. Extracted content material is processed by Onity’s customized AI mannequin. If the mannequin’s confidence rating meets the predetermined threshold, classification is full.
   3. If the doc isn’t acknowledged as a result of the mannequin isn’t skilled with that doc sort, the appliance mechanically routes the doc to Anthropic’s Claude Sonnet in Amazon Bedrock. This basis mannequin, together with different textual content and imaginative and prescient fashions akin to Anthropic’s Claude and Amazon Nova, can classify paperwork with out extra coaching, analyzing each textual content and pictures. This dual-model method, utilizing each Onity’s customized mannequin and the generative AI capabilities of Amazon, helps to optimally steadiness value effectivity with pace to market.
4. Extraction – Onity’s doc extraction software employs an algorithm-driven method that queries an inner database to retrieve particular extraction guidelines for every doc sort and knowledge factor. It then dynamically routes extraction duties between Amazon Textract and Amazon Bedrock FMs primarily based on the complexity of the content material.
   For instance, verifying notarization requires advanced visible and textual evaluation. In these instances, the appliance makes use of the capabilities of Amazon Bedrock superior textual content and imaginative and prescient fashions. The answer is constructed on the Amazon Bedrock API, which permits Onity to make use of completely different FMs that present the optimum steadiness of value and accuracy for every doc sort. This dynamic routing of extraction duties permits Onity to optimize the steadiness between value, efficiency, and accuracy.
5. Persistence – The extracted info is saved in a structured format in Onity’s operational databases and in a semi-structured format in Amazon S3 for additional downstream processing.

Safety overview

When processing delicate monetary paperwork, Onity implements sturdy knowledge safety measures. Information is encrypted at relaxation utilizing AWS Key Administration Service (AWS KMS) and in transit utilizing TLS protocols. Entry to knowledge is strictly managed utilizing AWS Id and Entry Administration (IAM) insurance policies. For architectural finest practices constructing monetary providers Trade (FSI) functions in AWS, confer with AWS Monetary Providers Trade Lens. This resolution is carried out utilizing AWS Safety finest observe steering utilizing Safety Pillar – AWS Nicely-Architected Framework. For AWS safety and compliance finest practices, confer with Greatest Practices for Safety, Id, & Compliance.

Remodeling doc processing with Amazon Bedrock: Pattern use instances

This part demonstrates how Onity makes use of Amazon Bedrock to automate the extraction of essential info from advanced mortgage servicing paperwork.

Deed of belief knowledge extraction

A deed of belief is a essential authorized doc that creates a safety curiosity in actual property. These paperwork are usually verbose, containing a number of pages of authorized textual content with essential info together with notarization particulars, authorized stamps, property descriptions, and rider attachments. The clever extraction resolution has decreased knowledge extraction prices by 50% whereas bettering general accuracy by 20% in comparison with the earlier OCR and AI/ML resolution.

Notarization info extraction

The next is a pattern of a notarized doc that mixes printed and handwritten textual content and a notary seal. The doc picture is handed to the appliance with a immediate to extract the next info: state, county, notary date, notary expiry date, presence of notary seal, individual signed earlier than notary, and notary public title. The immediate additionally instructs that if a area is manually crossed out or modified, the manually written or modified textual content ought to be used for that area within the output.

Instance output:

{
    "state": "Indiana",
    "county": "Monroe",
    "notary_date": "8/13/2024",
    "notary_expiry_date": "8/24/25",
    "notary_seal": "Current",
    "person_signed": "[Redacted]",
    "notary_public": "[Redacted]"
}

Extract rider info

The next picture is of a rider that features textual content and a collection of examine packing containers (chosen and unselected). The doc picture is handed to the appliance with a immediate to extract each checked riders and different riders listed on the doc in a offered JSON format.

Instance output:

{
"riders_checked": [],
"Others_listed": ["Manufactured Home Rider", "Manufactured Home Affidavit of Affixation"]
}

Automation of the guidelines overview of house appraisal paperwork

Residence appraisal reviews include detailed property comparisons and valuations that require cautious overview of a number of knowledge factors, together with room counts, sq. footage, and property options. Historically, this overview course of required handbook verification and cross-referencing, making it time-consuming and vulnerable to errors. The automated resolution now validates property comparisons and identifies potential discrepancies, considerably decreasing overview occasions whereas bettering accuracy by 65% over the handbook course of.

The next instance reveals a doc in a grid structure with rows and columns of data. The doc picture is handed to the appliance with a immediate to confirm if the room counts are similar throughout the topic and comparables within the appraisal report and if sq. footages are inside a specified share of the topic property’s sq. footage. The immediate additionally requests a proof of the evaluation outcomes. The applying then extracts the required info and gives detailed justification for its findings.

Instance output:

{
    "Outcome": "Sure",
    "Clarification": "Each situations are met. Room counts match at 4-2-2.0 (total-bedrooms-baths) throughout all properties. Topic property is 884 sq ft, and all comparable (884 sq ft, 884 sq ft, and 1000 sq ft) fall inside 15% variance vary (751.4-1016.6 sq ft). Comparable #3 at 1000 sq ft is inside acceptable 15% vary."
}

Automated credit score report evaluation

Credit score reviews are important paperwork in mortgage servicing that include essential borrower info from a number of credit score bureaus. These reviews arrive in various codecs with scattered info, making handbook knowledge extraction time-consuming and error-prone. The answer mechanically extracts and standardizes credit score scores and scoring fashions throughout completely different report codecs, reaching roughly 85% accuracy.

The next picture reveals a credit score report that mixes rows and columns with quantity and textual content values. The doc picture is handed to the appliance utilizing a immediate instructing it to extract the required info.

Instance output:

 {
    "EFX": {
        "Rating": 683,
        "ScoreModel": "Equifax Beacon 5.0"
    },
    "XPN": {
        "Rating": 688,
        "ScoreModel": "Experian Honest Isaac V2"
    },
    "TRU": {
        "Rating": 691,
        "ScoreModel": "FICO Danger Rating Traditional 04"
    }
}

Conclusion

Onity’s implementation of clever doc processing, powered by AWS generative AI providers, demonstrates how organizations can remodel advanced doc dealing with challenges into strategic benefits. Through the use of the generative AI capabilities of Amazon Bedrock, Onity achieved a outstanding 50% discount in doc extraction prices whereas bettering general accuracy by 20% in comparison with their earlier OCR and AI/ML resolution. The influence was much more dramatic in particular use instances—their credit score report processing achieved accuracy charges of as much as 85%—demonstrating the answer’s distinctive functionality in dealing with advanced, multiformat paperwork.

The versatile FM choice offered by Amazon Bedrock permits organizations to decide on and evolve their AI capabilities over time, serving to to strike the optimum steadiness between efficiency, accuracy, and price for every particular use case. The answer’s capacity to deal with advanced paperwork, together with verbose authorized paperwork, handwritten textual content, and notarized supplies, showcases the transformative potential of recent AI applied sciences in monetary providers. Past the quick advantages of value financial savings and improved accuracy, this implementation gives a blueprint for organizations searching for to modernize their doc processing operations whereas sustaining the agility to adapt to evolving enterprise wants. The success of this resolution proves that considerate software of AWS AI/ML and generative AI providers can ship tangible enterprise outcomes whereas positioning organizations for continued innovation in doc processing capabilities.

In case you have related doc processing challenges, we advocate beginning with Amazon Textract to judge if its core OCR and knowledge extraction capabilities meet your wants. For extra advanced use instances requiring superior contextual understanding and visible evaluation, use Amazon Bedrock textual content and imaginative and prescient basis fashions, akin to Amazon Nova Lite, Nova Professional, Anthropic’s Claude Sonnet, and Anthropic’s Claude. Utilizing an Amazon Bedrock mannequin playground, you may rapidly experiment with these multimodal fashions after which examine one of the best basis fashions throughout completely different metrics akin to accuracy, robustness, and price utilizing Amazon Bedrock mannequin analysis. Via this course of, you can also make knowledgeable choices about which mannequin gives one of the best steadiness of efficiency and cost-effectiveness on your particular use case.

In regards to the writer

Ramesh Eega is a World Accounts Options Architect primarily based out of Atlanta, GA. He’s keen about serving to clients all through their cloud journey.