The document discusses the journey from structured to unstructured data in insurance documents. It describes how omni:us uses computer vision, natural language processing, and machine learning to extract key information from insurance documents that may have highly variable layouts and formats. The presentation outlines how omni:us handles structured, semi-structured, and unstructured data through techniques like template alignment, text detection, and supervised learning to provide a scalable solution for insurance companies.
8. What we do
8
DOCUMENT CLASS
Policy
POLICY NUMBER
H 54/16 307 728
CUSTOMER
Renolate GmbH
10115 Berlin
AGENT
pma Insurance Broker
48149 Nurnberg
RISK DESCRIPTION / INSURED LOCATION
Private liability insurance comfort plus
Dog liability
Environmental damage insurance
Employees on premises
POLICY
Liability Protection
EFFECTIVE DATE OF CHANGE
22.12.2016 12:00
TERMINATION
22.12.2019 12:00
ANNUAL CHARGE
EUR 424,63
COVERAGES
Persons & property damage flat
Financial losses
Environmental damage basic flat
EUR 3.000.000
EUR 100.000
EUR 3.000.000
omni:us understands insurance documents
with highly variable layouts, and extracts
relevant data to radically streamline manual
processing of claims & policy comparisons.
omni:us Claim allows claim settling times
to be a matter of minutes. Not weeks.
omni:us Policy enables insurers to make
quicker, more accurate quotations &
comparisons.
omni:us @ datanatives 201822.11.2018
10. Who we are
Company
• Startup founded in 2015
• Berlin based with US subsidiary
• Enterprise customers in 6 countries on 2
continents
• Engineering and Scientific background in
CV/ML/AI
• 40+ team
10
USPs
• Fully scalable & global coverage
• Process-less integration and common
APIs
• Teachable AI
• Pretrained models for a variety of use
cases
nvidia Inception
Program
Red Herring
Winner 2017
Google
AI Program
Plug and Play
Insurtech
Member
omni:us @ datanatives 201822.11.2018
22. Real world challenges
• Various layouts
• Different versions
• Multiple languages
• Varying quality
• Diverse writings
• Light/shadows
• Misalignments
• Deformations
22omni:us @ datanatives 201822.11.2018
25. Template alignment
Objective
• Align existing form template to filled
pages
• Ease subsequent information extraction
Method
• Estimate global transformation
• Feature based approach for photos
• Dense pixel based approach for scans
Results
• Precise alignment for 95% of the pages
• Reliable despite drastic appearance
changes
25omni:us @ datanatives 201822.11.2018
27. Business benefits
27
• Same quality with 66% lower human time /costs
• Flexible deployment in the cloud or on premise
omni:us @ datanatives 201822.11.2018
41. omni:us @ datanatives 2018 41
Agenda
omni:us
Structured,
Semi-structured and
Unstructured data
Key-takeaways
22.11.2018
42. 42omni:us @ datanatives 2018
Key Takeaways
● Deep Learning to solve tasks and not datasets requires a very strong thought
process and engineering efforts
● Don’t have to be afraid of unstructured data anymore
● Solving a problem, requires multi-dimensional approach. No one tool can solve the
problem entirely.
22.11.2018