Implementing computer vision models for a digital microscope scanner with augmented reality features

Problem

• Product development of a digital microscope scanner with augmented reality features
• Delineations need to be created around areas of cancerogenous cells
• High resolution images (1px = 2 micron, 400 Mpix)
• Delineations ready below 5 secs

Solution

• Pixel-wise predictions
• Choice of hardware and software architecture for the inference

Impact

• AI contribution to the product development (TensorFlow, Keras, ConvNets, fully-convolutional network)

Defined requirements and helped selection of the best-fitting DWH platform to integrate the existing technical landscape

Problem

• 5 different business units leveraging a variety of on-prem and cloud data sources to drive business analytics capabilities
• Strong existing architecture landscape in place, based on AWS and Foundry
• Very diverse business needs across sectors, with strong preferences towards specific solutions to be mitigated

Solution

• Strong engagement with stakeholders across all BUs to drive requirements collection
• Assessment of 10+ candidates for DWH implementation
• Orchestrated procurement process with 4 different vendors to determine fit-for-purpose and fit-for-cost viability

Impact

• Selected the optimal DWH candidate to serve the company’s current and future analytics needs
• Ensured buy-in across all Business Units on the way forward to leverage the company’s data
• Enabled 1M$+ saving on selected platform cost vs other candidates

Built a capacity planning tool via Workshop on Foundry, providing visibility on the usage of equipment and enabling the simulation of “what-if” scenarios

Problem

• Capacity planning at a global life sciences company was a manual and tedious process
• The company needed better visibility on the current usage of equipment, and the ability to simulate the impact of changes

Solution

• Workshop application on Foundry
• Metrics monitoring and simulation capabilities via “what-if” scenarios
• Build up of a backend pipeline

Impact

• Reduction of manual work
• Quicker analysis of the impact of simulated changes
• Reduction of human error

Conducted a one-day workshop for a Swiss pharmaceutical company to advise them on the build-up of a Commercial Analytics CoE.

Problem

• Challenges in scaling analytics from PoC to enterprise level solutions
• Analytics team struggling to find more time for strategic advisory, overloaded with operational activities

Solution

• Knowledge transfer regarding organisational options & best practices for the set-up of an Analytics CoE
• Provided frameworks for the definition of the Analytics CoE’s scope and service levels
• Guidelines and framework  for the definition of the Analytics CoE’s operating model

Impact

• Identified the internal clients that the Analytics CoE must serve, along with their needs
• Defined the Analytics CoE’s value proposition and what is needed to enable it
• Drafted the high-level setup of the Analytics CoE

Developed an NLP pipeline for the automated classification of cross border invoices

Problem

• Wrong classification of cross-border invoices can lead to multi-million penalties risk on a yearly basis
• A global pharmaceutical company wanted to mitigate compliance risk related to the misclassification of cross-border invoices

Solution

• Developed an NLP pipeline, to automate the classification of cross-border invoices through an existing NLP model
• The NLP model reads the invoice and predicts the top classes with a corresponding confidence value
• Highest confidence code is then assigned to ERP field during the transaction posting

Impact

• Time to select coding is significantly reduced
• Increase of compliance and risk reduction

Problem

• Running out of production capacity, significant CapEx investment needed in the new production line to keep up with demand
• Goal: increase the production capacity by 5% via advanced analytics

Solution

• Data integration
• Anomaly detection algorithms
• Data correlations -> insights -> recommendations

Impact

• Trending towards 11% of productivity increase (throughput and yield optimization)
• 15x ROI

Designing and implementing a cloud-based data science platform based on AWS (EC2, SageMaker) with Foundry integration to streamline DS use cases

Challenge

• Lacking efficiency of data science projects due to insufficient computing resources
• Heterogeneous and distributed data sources, as well as low data quality
• Regulatory requirements with regard to data integrity

Solution

• Implemented data science platform based on AWS cloud infrastructure and services
• Ensured GxP compliance for regulatory audits of platform
• Established standard processes and procedures to manage and maintain the platform

Business impact

• Increased data science project efficiency due to lower data finding, connection, and preparation efforts
• Availability of modern cloud infrastructure based on AWS including integration with other platforms
• Reduced compliance and regulatory efforts (“compliant out-of-the-box”)
• 5 use cases went live in first 9 months on the platform

Designing new governance framework in collaboration with client team to streamline Foundry governance processes across the organisation

Challenge

• 4 semi-independent departments (data-wise) with only one using Foundry
• Current Foundry governance processes not scalable, new centralised “data office” (DO) to take-over organisation-wide governance efforts
• New data office had no Foundry experience

Solution

• Helped engineers in new team onboard on foundry
• Orchestrated upscaling of governance processes and coordinated with other DOs
• Designed new governance processes whenever necessary

Business impact

• Significantly improved governance processes with better access control
• All sector DOs now have access to necessary resources for use cases
• Lower request resolution waiting time thanks to more efficient governance framework

Designed and implemented advanced analytics platform on AWS allowing quicker launch of new projects, faster data access, and better project cost monitoring

Challenge

• Goal to help accelerate development of advanced analytics initiatives in the company
• Environment for quick and compliant project development necessary
• Need help designing and implementing central platform

Solution

• Designed and implemented a shared analytics platform on AWS
• Laid the foundational building blocks, i.e.,:

1.  Identity & Access Management
2.  Data Governance & Security
3.  Development Lifecycle
4.  Data ingestion, processing & presentation

Business impact

• Reduce time required to start new advanced analytics projects by implementing standard platform
• Reduce time needed to access the right data
• Allowed accurate monitoring of project costs
• 6 projects onboarded within first 3 months

Increased production line capacity by 11% through data analysis and ML modelling avoiding the need for significant CapEx investment to build new line

Challenge

• Manufacturing line running out of production capacity
• Significant CapEx investment needed to keep up with demand
• Goal to increase production by 5% through advanced analytics only

Solution

• Conducted exploratory data analysis to get good idea of production line and its overall equipment effectiveness (OEE)
• Prioritised most promising actionable levers to increase OEE with domain experts and simulations
• ML modelling, root-cause analysis, anomaly detection & optimisation
• Report detailing approach and gain for each lever

Business impact

• 22 improvement levers investigated in total; 6 retained and 3 implemented
• Improvements implemented led to faster overall production process
• 11% yield improvement
• 200 million extra tablets produced each year → CHF 40+mio in increased revenue
• 15x ROI

Conducted root cause analysis to uncover a production process irregularity problem, discovering a factor explaining 40% of production variability

Challenge

• Irregular production process causing missed deliveries and higher storage costs
• Goal is to use data-driven approach to investigate factors leading to instability in production process and find actionable solutions to be implemented

Solution

• Data/feature engineering and modeling implementation phases combined with tight collaboration with domain experts from differents areas of the org.
• Unified data  into one clean structure
• Conducted various statistical approaches to discover root cause

Business impact

• Root cause of dissolution problem not discovered, but found factor
  that explains 40% of variability in production
• Significant findings on what influences the production process and build solid datasets to enable deeper analysis in the future
• Delivered analyses and data documentation as well as user-training materials

Evaluating existing MLOps setups and tools used, providing improvement recommendations, and directly implementing some of those recommendations

Challenge

• Needed help in evaluating their MLOps setup that is currently being developed
• Mock data science project to be used by data scientists as an example to follow on future projects

Solution

• Evaluated existing tools and performed suitability assessment for the current and planned use cases
• Implemented some of our recommendations to provide clean split of responsibilities between the tools
• Recommended best practices on how to scale their infrastructure in the future

Business impact

• Significant improvements in existing MLOps setup by challenging status quo and providing recommendations
• Implemented best practices in their mock project example
• Praised by stakeholder for having also implemented things, going beyond mere recommendations

Designed and built a near real-time production process monitoring system to detect issues faster, enable proactive maintenance and monitor performance impact of improvements, all visualised on a dashboard

Challenge

• Lacked insights into what was happening on the production line
• No real-time sensor data analysis available
• Inability to detect issues on machines early on

Solution

• Designed and built a near real-time process monitoring system
• System collects data from manufacturing equipment and other sources and processes it to produce actionable insights
• Created dashboard to visualise information with an altering system on top

Business impact

• Better visibility into production process in near real-time
• Enabled faster root cause analysis
• Ability to monitor impact of improvements carried out on production line
• Proactive remediation to issues now possible
• Used daily by personnel

Developed a machine learning model that predicts the complexity of coding/tagging a stay given the clinical documents to improve billing correctness

Challenge

• Correct tagging of patient stay needed for accurate insurance processing
• Incorrectly tagged documents can be under-reimbursed or rejected
• Some documents are very complex to tag/classify, and others are routine
• Need to know if a document is somehow “off” (i.e. anomalous) from a content perspective

Solution

• Created data pipelines for clean aggregation of data
• Built application to automatically classify the “complexity” of documents, so they would be handled by the correct expert
• Same application created knowledge to enable working on anomaly detection

Business impact

• NLP model accuracy comparable with human expert coders
• Faster tagging for billing purposes
• More efficient workload distribution of coding tasks based on complexity and code skill
• Better decision support and minimised manual errors

Established a common development framework, as well as a standardised release process enabling faster and more efficient ML model deployment

Challenge

• Data Science team at client has been working on several ML projects that need to be put into production
• Goal of this project was to provide an initial MLOps process to deploy models into production

Solution

• Consulting and implementation of right technology stack combined with establishing a proper process
• Established common framework (Kedro) to develop ML models. Setup a standard release process to deploy ML use cases
• Defined tech stack needed for future use cases

Business impact

• Established common development framework to standardise processes
• Best practices put in place for development and release of software/ML models
• Built initial infrastructure enabling reproducible ML pipeline creation on-premise

Building interactive dashboards on PowerBI from telemetry and sales data enabling salesforce and management to derive actionable market and customer insights

Challenge

• Telemetry and sales data have been collected separately but were not linked to deliver business value
• Management and sales representatives were not able to derive business insights as data could not be visualized in real time

Solution

• Built an interactive PowerBI dashboard, aggregating data from various data sources
• Included “smart insight recommendations” based on real time data
• Implemented various filtering options to analyze KPIs based on multiple criteria

Business impact

• Improved decision-making through inclusion of real time data to derive market insights
• Increased sales efforts enabled by better client-specific recommendations
• Generated high level insights allowing management to see things under a different lights

Problem

• Difficulties to identify the patients eligible for the particular treatment, based on narrow criteria in the free text patient records

Solution

• PoC to create a way to look up patients meeting the particular criteria using pattern matching and machine learning so that they can be qualified for treatment

Impact

• The eligible patients will receive the treatment they are eligible for
• Time saving compared to a manual lookup

Problem

• Pneumonia causes over 1M deaths a year
• Cost effective process needed for diagnosing X-Ray scan images

Solution

• Pro-bono research work to develop a predictive model for the diagnosis

Impact

• 92% of accuracy
• Model published -> available for researchers

Problem

• The customer’s production line went through a data driven optimization, the customer wants to ensure the production line maintains the highest possible OEE and wants to add advanced features like automated issue detection & prediction

Solution

• Near real time continuous monitoring system of key production processes with advanced analytics built in (alerting, trend detection, anomaly detection, predictive maintenance)

Impact

• Greater production process visibility
• Quicker time to resolution
• Improved uptime

Problem

• Running out of production capacity, significant CapEx investment needed in the new production line to keep up with demand
• Goal: increase the production capacity by 5% via advanced analytics

Solution

• Data integration, AWS platform
• Anomaly detection algorithms
• Data correlations -> insights -> recommendations

Impact

• Trending towards 11% of productivity increase (throughput and yield optimization)
• 15x ROI