How PellencST cut their time-to-model by 90% and gained a competitive edge in the sorting machine market
Transforming waste sorting machines with vision AI
Unleash growth potential
Document Processing
How to maintain industry-leading product performance with the same staff as workloads and complexity increase dramatically, without compromising personal data security.
DOCUMENT PROCESSING
CTO
PRODUCTION
Context
A growing remote-identity-verification scale-up is expanding into both new geographies and new industries. This impacts significantly the complexity and the amount of workload for the machine learning and the data science teams.
Each country has a variety of document types (passport, driver's license, ID card, utility bill, professional license, etc.) and each document type has a number of versions that can be supported. The set of required documents changes according to use cases (financial services, mobility, healthcare, gaming, etc.) and new documents are issued regularly.
4 types of Computer Vision analysis are performed on each document: Detection, Segmentation, Classification and Tampering.
why choose us
4 weeks
From 3 months to 4 weeks to deploy a new tampering model in production.
2 weeks
From 4 months to 2 weeks to support a new document in production.
Double
The number of countries and user-cases addressed.
Challenge
1
Manage increasing workloads and complexity with the same number of people.
2
Continually add new document types and versions to the datasets.
3
Keep up with customer demands and fraudsters by training and deploying updated models.
4
Maintain ownership of personal data and ensure its security.
Solution
Tagging System: Organize data efficiently by tagging documents by type, country, format, security features, etc.
Dataset versioning: Enrich datasets with new documents and human-reviewed production images. Track dataset evolution.
Continuous training & deployment: Automatize model improvement and deployment.
Hybrid deployment: Store data and deploy models on their own infrastructure. Leverage the Picsellia GDPR infrastructure for model training.
FR
EN