Causal Inference Library
Multinational Consumer Electronics Company
Developing internal causal inference library consisting of 10+ algorithms and applied library to 2 use cases
Challenge
- Build causal inference library to be used by other data scientists in the future for various projects
- Enable activities such as churn analysis and A/B testing to be conducted
Solution
- Developed library in collaboration with internal stakeholders, did unite testing/documentation/user guide
- Worked on two related use cases needed for the analysis
- Unit8 implemented 60% of the algorithms for the project
Business impact
- Library enables data scientists to easily use multiple algorithms for causal inference through a unified API design
- 10+ algorithms implemented
Due Diligence Tool
US Law Firm
Building an application to streamline the due diligence process through a combination of automated web searches and web scraping
Challenge
- Client offers due diligence as a service
- Due diligence process is multi-tenant; users from multiple separate parties work together
- Process slowed down by back-and-forth communication and repetitive tasks
- Goal was to automate due diligence process
Solution
- Built a multi-tenant application to streamline the due diligence process end-to-end
- Implemented automated web search, website scraping and screening of entities linked to the target company
Business impact
- Project still in deployment, no impact available at this time
Digital Twin & Production Line Simulation
Glass Bottle Manufacturer
Developing tools on Foundry and applying causal models to identify optimal steps in glass bottle production process to increase yield of production line
Challenge
- Want to improve yield on production line form 86% to 90%
- Already identified step in process where gain could be realised
- But no way to automatically prioritise sensor data which should be further analysed
- Goal is to apply ML models to identify most promising captors to improve yield
Solution
- Initially built E2E model of yield on production line, but hit data quality roadblock
- Shifted towards:
- developing tool on Foundry to surface “golden production parameters” when production quality was high
- using causal models to understand optimal actions to stabilise first step of production process
Business impact
- Highlighted data quality issues and suggested next best steps for improvements
- Demonstrate Foundry’s ability to replace existing system
- Trained key internal stakeholders on Foundry platform
- Due to lack of high quality sensor data → no direct impact on yield realised during the PoC
Supply Chain Demand Forecasting
Medical Equipment Manufacturer
Building a forecasting model for a medical equipment manufacturer resulting in significant forecasting accuracy improvements allowing better procurement planning and contract management
Challenge
- Forecasting challenges when it comes to their supply chain
- Demand Planners working very manually with back and forth between planners, sales, and manufacturing
- Goal is to improve this with a more intelligent tool
Solution
- Built forecasting model with features included in Darts in combination with hierarchical reconciliation
- Solution would stream data directly from SAP
Impact
- Forecasting improvements of 10-60% compared with historical forecasting methods, depending on the time outlook in the model
- Helps procurement negotiate better contracts with suppliers and plan manufacturing accordingly
Data Management, Forecast, and Visualization on Foundry
Swiss Raw Material Vendor
Demonstrated how a cloud platform can ingest different data sources and made dashboards and applications that could drive strategic decision-making
Challenge
- Client wanted to have a view of the current raw material inventory in the warehouse
- Need to forecast the inventory on customer side in order to optimize production and logistic cost and to increase customer satisfaction
Solution
- Data ingestion from different data sources including SAP, MeteoSwiss, and IoT devices
- Using Supply Chain offerings to define ontology
- Clean data
- Apply forecasting models through our Darts library
Business impact
- Showcase different forecasting models alternatives
- Optimized warehouse logistics
- Reduced excess inventory stock
Forest Wildfire Spread Prediction
World Wildlife Fund (WWF)
Developed a wildfire prediction system in collaboration with WWF to anticipate wildfire in Bolivian forests through a new fuel model based on geo data
Challenge
- Bolivian forests are at a high risk of wildfire
- Wildfire prediction system needed to anticipate how a fire might start and where it will spread
Solution
- Used publicly available data such as weather, climate, past fires, and geo data to model the wildfire spread starting at a certain point in time
- Created a new fuel model for Bolivia by transferring knowledge from available US data
Business impact
- A complete fuel model for the Bolivian territory is now available
- Applications created enable visualisation of past fires
- Discovered main pain points of ML with geo data and established a clear plan on how to tackle them and speed up future use cases