## Overview FiscalNote is a software company founded in 2013 that specializes in transforming access to legislative and regulatory data at local, state, and federal levels. The company has developed machine learning (ML) and natural language processing (NLP) technologies to provide policy and geopolitical insights to a diverse customer base that includes nonprofits, government agencies, and over half of the Fortune 100 companies. Their platform offers automated analysis of legislative outcomes, policymaker effectiveness, and sentiment analysis to help stakeholders engage with complex policymaking processes. This case study focuses on how FiscalNote addressed significant bottlenecks in their AI/ML model deployment processes by adopting Databricks' unified data analytics platform, ultimately achieving 3x faster model deployment and improved analyst productivity. ## The Challenge: Fragmented ML Model Deployment Infrastructure Before adopting Databricks, FiscalNote faced substantial hurdles in deploying and maintaining their AI/ML models. The company works with both structured and unstructured data sources to provide legislative intelligence, which requires sophisticated NLP and ML capabilities. However, their existing infrastructure created several pain points that impacted their ability to serve clients effectively. The primary challenge was the cumbersome, fragmented process required for model deployment. Data teams had to manually stitch together various components essential for the ML lifecycle, including tracking artifacts, managing builds, and maintaining model versions across different notebooks. This extensive manual process severely limited the team's deployment cadence—they could only deploy or update certain AI models approximately once per year. For a company operating in the rapidly evolving legislative landscape, this annual deployment cycle was a significant competitive disadvantage. As CTO Vlad Eidelman explained, clients required models in different locations, of varying sizes, and with different capabilities to satisfy their accuracy needs while maintaining operational efficiency. Each deployment phase extended over several weeks, which restricted FiscalNote's ability to respond quickly to legislative changes and provide up-to-date models to clients who depend on timely intelligence. A second major challenge involved the requirement for zero-disruption deployments. FiscalNote's infrastructure demanded extensive custom coding to ensure that model updates would not interrupt the operation of existing models. This approach increased the operational burden on team members and introduced risks to data integrity and service continuity. The team spent considerable time and resources on defensive engineering rather than advancing their ML capabilities. Additionally, a lack of discoverability of essential data assets hindered effective model development. Data scientists struggled to access and utilize necessary data, which impeded the entire model development and deployment pipeline. This created an environment where operational overhead consumed resources that could otherwise be dedicated to innovation and model improvement. ## The Solution: MLflow and Mosaic AI Model Serving FiscalNote chose Databricks as their unified data analytics platform to address these challenges, focusing on two key components: MLflow and Mosaic AI Model Serving. ### MLflow for ML Lifecycle Management FiscalNote's initial implementation focused on MLflow, the open-source platform developed by Databricks for managing and deploying ML models. The primary objectives were achieving enterprise scalability and reliability while reducing the operational burden on data science teams. MLflow enabled the data science team to automate the tracking of artifacts, model versions, and notebooks—tasks that had previously consumed significant engineering time. This automation freed data scientists to focus on tasks that genuinely required human expertise and couldn't be handled by automated systems. The platform provided a centralized registry for models and experiments, eliminating the fragmentation that had previously plagued their deployment processes. The adoption of MLflow translated directly to operational improvements. FiscalNote experienced faster time to market and gained the ability to scale their model deployments on an as-needed basis rather than being constrained to annual release cycles. The reduction in MLOps overhead allowed the team to invest more heavily in data science innovation. ### Mosaic AI Model Serving for Streamlined Deployment FiscalNote also leveraged Mosaic AI Model Serving, a comprehensive service for deploying, managing, and monitoring machine learning models. This component addressed several critical pain points in their deployment workflow. With Model Serving, FiscalNote's data teams no longer needed to worry about building APIs, hosting infrastructure, scaling considerations, or planning zero-disruption deployments. The service abstracts away these operational concerns, allowing teams to deploy models without the extensive custom infrastructure work previously required. The Model Serving capability enabled FiscalNote to roll out various ML models tailored to specific use cases. The text mentions several specific applications: - Supporting standard ETL pipelines to ingest data from various sources - Performing NLP tasks such as document summarization and sentiment analysis - Executing binary classification workflows, including analyzing whether a congressperson will or will not vote on a particular bill According to Eidelman, the adoption of Databricks Model Serving began within the six months prior to the case study publication. Before this adoption, teams had to manually store, serve, and build models while piecing together all the required infrastructure. The shift allowed them to focus more on approach and methodology rather than logistics and infrastructure management. ## Results and Business Impact The adoption of Databricks tools delivered measurable improvements to FiscalNote's operations: **3x Faster Model Deployment**: The most significant quantified result was a threefold increase in the number of models deployed annually compared to their pre-Databricks output. This acceleration directly correlates with faster time to market for new analytical capabilities. **Increased Analyst Productivity**: By automating MLOps tasks and eliminating infrastructure management overhead, data scientists and analysts became more productive and effective. They could dedicate their expertise to model development and improvement rather than operational tasks. **Enhanced Responsiveness to Legislative Changes**: The faster deployment capability enables FiscalNote to be more dynamic in responding to legislative changes. This agility allows them to process and analyze legislative data quickly and accurately, delivering timely intelligence to clients who need to adjust to new laws and regulations. **Freedom to Experiment**: The efficiency gains empowered the data science team to experiment freely and iterate rapidly on models. This capability ensures that solutions remain aligned with current market needs and client requirements. ## Critical Assessment While the case study presents compelling results, several points warrant consideration: The case study is published by Databricks, the vendor providing the solution, which introduces potential bias in the presentation of benefits. The 3x improvement in model deployment is significant, but the baseline (annual deployments) was notably constrained, suggesting the previous infrastructure may have been particularly inefficient. The case study focuses primarily on traditional ML models (classification, NLP tasks) rather than large language models specifically. While NLP and sentiment analysis are mentioned, there's limited detail on whether FiscalNote is leveraging foundation models, fine-tuning approaches, or primarily using classical ML techniques. The "LLMOps" framing is somewhat generous—this appears more accurately to be an MLOps transformation that could support LLM deployment in the future. The mention that Model Serving adoption began "within the last six months" suggests the full impact of this component may not yet be fully realized or measured. Long-term sustainability of these improvements would require continued evaluation. ## Technical Architecture Notes The case study mentions AWS as the cloud provider, placing FiscalNote's Databricks deployment in the AWS ecosystem. The architecture combines: - **MLflow**: For experiment tracking, model versioning, and artifact management - **Mosaic AI Model Serving**: For model deployment, API management, scaling, and monitoring - **ETL Pipelines**: For data ingestion from various structured and unstructured sources - **NLP Workloads**: Including summarization, sentiment analysis, and classification tasks This architecture represents a modern MLOps approach that centralizes ML lifecycle management while providing the flexibility to deploy models for diverse use cases across FiscalNote's product offerings serving Fortune 100 companies, government agencies, and nonprofits.