Choosing the right IDP Solution

Choosing the Right Document
Processing Solution
Presented by:
Almir Davletov, IDP Subject Matter Expert @ Provectus
Yaroslav Tarasyuk, Business Development @ Provectus
Sonali Sahu, Sr. Solutions Architect @ AWS

1. Provide an overview of the market for document processing solutions
2. Outline critical factors for choosing the right document processing solution
for your business use case
1. Strategize on whether you should look for a ready-made solution to purchase,
or to build a custom solution of your own
1. Get qualified for the Provectus IDP Solution Discovery Program
Webinar Objectives

Almir Davletov
IDP Subject Matter Expert,
Provectus
Sonali Sahu
IDP AI/ML Specialist,
Healthcare & Life Sciences,
AWS
Yaroslav Tarasyuk
Business Development,
Provectus
Introductions

Major Business Drivers
Companies are looking to adopt IDP solutions to unlock insights from semi-structured and unstructured
documents to improve operational efficiency and strategic outcomes.
Major business drivers of IDP adoption are:
1. Reduce costs of document processing
2. Increase accuracy and speed of processing
3. Improve strategic business outcomes by providing actionable insights for increased customer
satisfaction & employee productivity
4. Potential synergy by integrating IDP solutions with RPA or BPM solutions

Technology Consulting
& Professional Services
650+ Employees
● US, Canada, LATAM, Europe
Competencies
● AWS Premier Partner
● AI/ML, Big Data & Analytics
● DevOps, Cloud Migration
Industry Focus
● Healthcare & Life
Sciences
● CPG, Retail & Ecommerce
● Manufacturing
● Public Sector
AI Solutions
● Crystal: Customer 360 Personalization
● IDP: Intelligent Document Processing
● Real-Time Video Analytics
● Fraud Detection
Foundation & Accelerators
● NextGen Data Lake
● MLOps Infrastructure
Open-Source Leadership
● Kubernetes, Kubeflow
● UI for Apache Kafka
● ODD: Open Data Discovery
Spec & Platform
Provectus’ mission is to leverage cloud, data and AI to reimagine the way
businesses operate, compete, and deliver customer value

Documents are everywhere
1 2 3 4
structured semi-structured unstructured handwritten
Legal
● Contracts
Banking & Finance
● Bank & income
statements
● Insurance claims
● Invoices & contracts
Healthcare
● Patient onboarding
● Medical records
● Claim-related
documents
Supply Chain
● Shipping labels
● Proof of delivery
● Bill of lading
Manufacturing
● Purchase orders
● Change requests
● QA records
Human Resources
● Employee
onboarding
Government
● Immigration
applications
● Tax assessment
forms
Telecoms
● Maintenance logs
● Driver logs

OCR converts images of documents into plain
text and extracts specific fields based on
templates, regular expressions, etc.
It uses rule-based or template-based
extraction. User needs to configure the
templates for each document type
Every converted document needs to be
manually reviewed, unless the input documents
are standard (in quality, positional elements,
etc.)
Cannot process unstructured documents such
as contracts and emails
It may use OCR to convert images of documents
to a digital format, but extracts specific
information using machine or deep learning
algorithms
The extraction does not depend on the template
but content. User needs to do minimal (if any)
training for minor template changes
Once the system is trained, Straight Through
Processing (STP) can be partially enabled. Human
interaction is reduced to minimum
Using Natural Language Processing (NLP) models,
IDP can extract specific information from complex
unstructured documents
OCR and template-based solutions IDP Solutions
Compare OCR and IDP

General goal is to spot main entities
in the document (paragraphs, forms,
tables, etc.) and then successfully
identify written text in them
(segmentation and OCR).
IDP Components: OCR/CV

Context search on data from OCR + segmentation
Forms and tables greatly impact overall performance. Data extraction from forms is resolved
efficiently (due to a straightforward key-value structure), but tables are still being a pain point
for all the data extractors. For unstructured texts, deep networks are a solution at this point. E.g.
ReLIE or GPT-like models are good for finding key-value (question / answer) pairs in the context
of the unstructured documents.
OCR +
Segmentation
Users input
Semi-structured
data
Query
Context
search
Structured
answers
IDP Components: Data Extraction

Evaluation of the document processing
model is a task in progress.
Results with a low-confidence score
and missing information
are forwarded to human experts.
Samples of successfully extracted
information are also forwarded to
human experts for evaluation.
IDP Components: Evaluation and Monitoring

S3, RDS, ElasticSearch
Input documents are usually stored as is in
a Data Lakes, application data and metadata -
in RDS, discoverable output results -
in ElasticSearch.
But depending on the industry, storage layer
can get more complicated. For example, FHIR
data format in the healthcare industry requires
careful design and implementation of the data
storage.
IDP Components: Storage

Automation encapsulates all the processes outlined previously
and unites them into one single product, featuring:
● Document capture
● ML model lifecycle
○ Labeling
○ (Re)Training
○ Evaluation
○ Monitoring
● Human-in-the-loop
● Integrations
● CI/CD pipelines
● System monitoring
● Infrastructure as Сode
IDP Components: Automation

IDP is more than just OCR. To resolve the problem in-house, you need
to take care of data capture, data ingestion, pre-processing, OCR, data
extraction, evaluation, post-processing, and further integrations to
destination systems.
Bottleneck: Tables and unstructured text
IDP Takeaway

Solutions Landscape
Market Overview

Public Cloud
Solutions by public cloud providers are
built on top of fully managed services
and easy to setup and deploy.
1. Maintained by cloud vendor
2. Built-in infrastructure
3. Secure and compliant
4. Human-in-the-loop included
5. Low costs at scale
6. Mostly general-purpose, but
with few specific use cases
Standalone
Pre-packaged solutions by ISVs,
usually narrowed down to a several
use-case and shipped as SaaS.
1. Little or no configuration
2. Narrow use-case support
3. Usually comes with rich GUI
4. Human-in-the-loop included
5. Medium or high costs
6. Data usually stored outside
of the organization
Custom In-House
Solutions built in-house, implementing
very specific or custom use cases, can
be deployed on the cloud or on-prem.
1. Highly customizable
2. Room for modern technologies
3. IP belongs to the organization
4. Can onboard new use cases
5. High costs on the start
6. Requires strong domain and
technical expertise
Brief Solutions Overview

Public Cloud Vendors usually offer general-purpose technology components for document processing,
such as Amazon Textract, Comprehend, A2I, etc. Components like A2I has rich spectrum of functionality
allowing to handle edge cases using human forces.
Some services are narrowed down to the industry-specific use cases to improve the accuracy of
predictions, like Comprehend Medical, Textract Expenses Analyzer, Textract Identity Documents Analyzer,
etc.
Public Cloud Solutions

Stand-alone Solutions
Independent Software Vendors who have built solutions using AI-native platforms tackle the most
demanding automation challenges.
Generally, they can handle documents that are more complex or have greater variation. As a result, they
often can deliver a greater business impact for the narrow use cases. But the downside of such solutions is
cost and flexibility to onboard new use cases.
It also makes sense to assess the effort to integrate those solutions with the existing infrastructure and
systems.

Custom In-house Solutions
Custom in-house IDP solutions allow to implement most complex niche use cases by combining the
capabilities of the cloud vendors and modern state-of-the-art technologies. Full control over the roadmap
allows to build next-gen, future oriented solutions.
It requires strong technical and domain expertise, which a lot of the companies are lacking today, but
partnering with the IDP partners might efficiently fill this gap.

© 2022, Amazon Web Services, Inc. or its affiliates. All rights reserved. Amazon Confidential and
Trademark.
Amazon enables a fully
automated IDP
workflow, no ML
experience required
Benefits of Amazon’s Intelligent
Document Processing (IDP)
Go beyond OCR with
accurate, versatile
information extraction
Serve end
customers faster
Analyze documents and send
key insights to downstream
systems and workflows
Reduce the total cost of
document processing
19

Trademark.
Intelligent Document Processing pipeline
Documents
Categorized
Documents
Convert
Image to
Text
Data capture
S3
Bucket
Amazon
Comprehend
Classification
Amazon
Comprehend,
Medical
Amazon
Textract
Post processing
completeness
check, rule
based validation
& review
Model tuned
using labeled
data provided
by customer
Amazon
Textract
Amazon A2I
Route based on
business rules /
type of info
needed
Forms
Tables
IDs
Expenses
Find insights in unstructured
text
Identify Person, Places, and
business specific terms
Identify PII & redaction
Route High
confidence
predictions Send Data to
Downstream
Databases/Apps
Route Low
confidence
predictions
standard process
Categorize, Tag documents
with Classification
Verification &
Human Review
Get key insights with
Extraction & Enrich

Trademark.
Trademark.
Services used in the pipeline

Trademark.
Aggregating and storing documents
Storing documents in a
central location, such as
Amazon S3, for each
business use-case, so
they are ready for
processing
Scans from
physical mailrooms
Docs from
digital mailrooms
Uploaded by end
user using a
computer or phone
Email
Attachments
Scanned documents
PDF, JPEG, PNG, and TIFF
Captured by
cell phone cameras or digital
scanners
Files in common
formats such as Word
or .TXT

Trademark.
Identify document types
Sort different types of documents into the right category with
Amazon Comprehend Custom Classifier, and send them to the
correct document pipeline.
To train a custom classifier you need to provide 50 samples of each type of
document to teach the IDP AI how to classify your documents.
Real-time or asynchronous processing
Perform real time document classification when needed or reduce
costs using asynchronous processes.
Document classification
Contracts Enrollment
Forms
Provider
Invoices
Amazon
Comprehend
PD
F
TIF
F
PN
G
JP
G
Amazon
Textract
Converts to text
Tex
t
Pathology
Reports
Visit
Summaries

Trademark.
IDP data extraction using Textract
Handwriting Tables
24
Text Specialized documents
Forms
Loan application
Invoices
and receipts
Identity
documents
Vendo
r

Trademark.
IDP extraction & enrichment with Amazon Comprehend
Discover insights and relationships in text
Key phrases
Language
Topics
Entities
and custom entities
PII
Personally Identifiable
Information
Automatically extract
insights from text
Amazon
Comprehend
Amazon
Textract
Documents
Financial, legal,
contract, ID, claims
documents

Trademark.
IDP enrichment with Amazon Comprehend
26
Redaction of
sensitive data
DOCUMENT TEXT
Hi, my name is John Doe.
For verification, the last 4
digits of my social are 6789
and my DOB is 01/01.
Summary fields:
Category of entity:
Policy_ID
Entity: 456-YQT
Confidence:
0.95
Recognize entities
specific to your business
DOCUMENT TEXT
Hi my name is John. I am filing a claim
for a car accident, my insurance
number is 456-YQT.
OUTPUT
Hi, my name is
[NAME]. For
verification, the
last 4 digits of my
social are [SSN]
and my DOB is
[DATE_TIME].
OUTPUT
Organization: Budget Mutual
insurance Company
Location: 9876 Infinity Ave
Springfield, MI 65541
Date: 6/15/20202
Recognize
common entities
DOCUMENT TEXT
Budget Mutual insurance
Company 9876 Infinity Ave
Springfield, MI 65541
Client Laura McDaniel and date
of death is 6/15/20202
Category of Entity:
MedCondition
Type: Dx name
Entity: Hypertension
Confidence: 0.99
Recognize medical
entities
DOCUMENT TEXT
Mr. Smith is a 63-year-old
gentleman with coronary artery
disease and hypertension.
CURRENT MEDICATIONS:
taking a dose of LIPITOR 20
mg once daily.
OUTPUT

Trademark.
Use human in the loop (A2I) when needed
27
Human
Review
Business rules drive when to
review
• Example: review if low field accuracy or if
claim value is >[$25K]
• Flexible to use your staff or your
designated 3rd party
• Can use multiple reviewers for most
critical data points
• Admin panels allow business users to
update rules

● Provectus IDP Platform is a white box solution integrated with the
expert-in-the-loop UI, retraining pipelines, analytics dashboards,
built-in integrations and custom built NLP models that go beyond
the generic use cases covered by the public cloud vendors and
standalone solutions.
● It can be deployed on top of existing services like AWS Textract and
Comprehend to combine the best from two worlds - fully managed
OCR engine and custom built CV/NLP models.
● The “Platform” nature of the Provectus IDP solution allows to
customize it in every step, covering all the business end-to-end and
allowing to seamlessly integrate with any existing systems and
infrastructure.
Solution Overview

Structured &
unstructured data
Including invoices, contracts,
insurance claims, shipment
orders
Provectus IDP
White box solution that
processes data
integrated with Cloud
OCR, expert-in-the-loop
UI, retraining pipeline,
custom built NLP models
Structured, indexed,
and normalized data
Including invoices,
contracts, insurance claims,
shipment orders, medical
records, etc.
OCR, Data
Extraction
CV/NLP models
Pre-Processing
Classification
APIs, Integrations
Data Validation
Expert-in-the-loop
UI and retraining
Amazon
SageMaker
Amazon
QuickSight
Third-party
applications
How It Works on AWS

● Provectus IDP is a platform, allowing to either start with the pre-
configured deployment for industry-specific use cases or implement
complex and specific use cases
● We train-retrain our models for each customer, so the quality of our
models stacked on top of Textract allow them to close the gap and
outperform competitors on the given range of the documents
● Human-in-the-loop system that allows to review disputed cases
and input corrected data into the ML model to increase text
extraction accuracy
● Support of structured, semi-structured and unstructured
documents
● Ensuring highest security and compliance standards
Key Features

Accurate Extraction and Conversion
Capable of ensuring 80-95% accuracy of data extraction,
the solution flags disputed cases for manual review to
prevent any errors in the output data.
Continuously Improved Accuracy
Powered by a Natural Language Processing technologies,
solution continuously retrains the ML models and improves
its self-sufficiency.
Streamlining Business Processes
As human review is needed only for a error correction and
data verification, companies can scale their document
processing workflows and reduce the costs.
Faster Access to Analytics Initiatives
Scanned documents are processed at scale, and
businesses get faster access to structured data to drive AI
adoption and analytics initiatives.
Lower document processing costs and error-rate
Key Benefits

From 4-8 months of internal
R&D to free of charge 4 week
PoC
Accelerate your time
to market
10-30% of quality improvement
on documents processed by
pure cloud components or
standalone solution vendors
Push the limits of cloud
and stand-alone solutions
● By 2-8x compared to manual
workflows
● By 30%+ compared to legacy
OCR solutions
● By 10%+ compared to modern
cloud solutions
Reduce your document
processing costs
Understanding the Value

Vendors Evaluation
Methodology

Data
● EDA (exploratory data analysis) —knowing your data is
the key to success
● Sample data based on EDA
● Use this data as the evaluation dataset for measuring
performance of solutions on the market / in the segment
Metrics
● F1, accuracy, recall, etc.
● Key / value extraction
● Table data
● Language, character recognition, spelling, handwritten text
Provectus Evaluation Methodology

Now, you have all the information about
possible go-to solutions in your market
segment. What’s next?
You need to fairly compare each and every
solution to choose one that fits
and aligns with your use case the most.
Deep evaluation is key to making
the right decision.
Cloud-based
Stand-alone
Custom In-house
Provectus IDP
Platform on AWS
Amazon IDP
(Augmented AI)
What to Choose?

1. Ecosystem matters: Data integration with built-in industry specific connectors, data
pipelines, OCR, NLP, security, storage, and a human-in-the-loop workflow — all these
elements should be integrated with each other for optimal performance.
1. Use unbiased benchmarking framework for evaluating real performance of different
providers, based on your use case and datasets.
1. Work with Provectus to reduce your document processing costs
a. By 2-8x compared to manual workflows
b. By 30%+ compared to legacy OCR solutions
c. By 10%+ compared to modern cloud solutions.
Takeaways

Questions?
We would be happy to answer!
125 University Avenue
Suite 295, Palo Alto
California, 94301
provectus.com

Choosing the right IDP Solution

More Related Content

Similar to Choosing the right IDP Solution

More from Provectus

Recently uploaded

Choosing the right IDP Solution