Inspiration
We love the idea that data-driven decision-making doesn't have to be dull! When we discovered Federato's RiskOps, a cool SaaS platform that helps insurance underwriters navigate complex business risks, we knew we had to get involved. Our inspiration came from wanting to use the power of data science to make users' experiences not only efficient but enjoyable, creating insights that actually matter.
What it does
Our awesome data pipeline transforms vast amounts of event-based user interaction data into clear, actionable insights. It reveals patterns that help keep users hooked, recommending personalized next steps in real-time to make sure underwriters stay engaged, productive, and happily exploring high-value features.
How we built it
We developed a comprehensive analytics pipeline utilizing Python alongside advanced analytical tools and frameworks:
- Data Processing & Management: NumPy, Pandas, Joblib, PySpark
- Interactive Visualization: Plotly, Streamlit
- Advanced Predictive Modeling: XGBoost, Prophet, ARIMA, KMeans, LSTM, Markov Chain Models
- Performance Optimization: Chunk-based processing, CSV-based intermediate storage
Key steps included:
- Column Cleanup: Removed irrelevant columns, retaining only features critical for analyzing user engagement.
- Geographic Data Imputation: Implemented KNN imputation to handle extensive missing location data.
- Filtering & Feature Engineering: Focused on relevant user segments, excluded incomplete data, and introduced vital time-based features such as session durations and latency metrics.
- Outlier Handling: Utilized z-score methodologies to remove data anomalies, ensuring accuracy and reliability.
- Predictive Modeling: Trained advanced predictive models (XGBoost, Markov Chains, and LSTM) to analyze sequences of user actions and their impact on retention. These models assessed user moves by identifying critical interactions and transitions that significantly enhance engagement and session duration.
Challenges we ran into
- Handling and analyzing a large volume of data efficiently (1.85 million events in 2025, 14 million+ events after preprocessing for 2024)
- Managing substantial missing geographic and temporal data
- Ensuring precise synchronization of client-server timestamps
- Balancing detailed data cleaning with optimal computational performance
We overcame these challenges by employing parallelized processing, strategic chunking, and iterative enhancements in imputation and feature engineering.
Accomplishments that we're proud of
- Cracking open user behavior data to find insights that truly matter
- Building accurate, fun-to-use predictive models for recommending user actions
- Creating a real-time optimization framework using advanced analytics like Markov Chains and XGBoost
- Crafting engaging visualizations that clearly map out user journeys and interaction trends
What we learned
Through the Federato RiskOps Optimization Challenge, we gained valuable insights into:
- Real-world SaaS data can be messy, especially in insurance—but it’s super rewarding to tame it!
- Solid data preprocessing is the unsung hero behind every powerful predictive model
- Causal inference isn’t just academic—it helps uncover real, impactful user actions
- Efficient processing and quality data don’t have to be mutually exclusive
What's next for Maple Valley
Moving forward, we aim to:
- Embed our optimization recommendations right into Federato's RiskOps platform for instant, real-time engagement boosts
- Keep improving our predictive models with fresh data and new insights
- Dive even deeper into causal analyses to reveal hidden gems about user retention
- Expand our analytics toolkit to help even more SaaS platforms across the insurance and risk management landscape
Log in or sign up for Devpost to join the conversation.