Streamlining Legislative Analysis Model Deployment with MLOps

fiscalnote · Legal · 2024 · source Agentic: no · promotional bias: high · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Legislative and regulatory intelligence model deployment

FiscalNote needed to deploy and update ML/NLP models faster for legislative and regulatory intelligence, but fragmented MLOps infrastructure forced manual artifact, build, model-version, API, hosting, scaling, and no-disruption deployment work.

Constraints:

Architecture

FiscalNote adopted Databricks MLflow and Databricks Model Serving. MLflow manages artifact, model-version, notebook, and deployment lifecycle tracking so data scientists spend less time on MLOps logistics. Model Serving deploys, hosts, scales, manages, and monitors ML models without custom APIs or infrastructure, supporting ETL, NLP summarization/sentiment, and binary classification workflows for legislative intelligence.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  legislative_and_regulatory_data["Legislative and regulatory data"] --> data_science_model_development["Data science model development"]
  data_science_model_development --> mlflow_lifecycle_management["MLflow lifecycle management"]
  mlflow_lifecycle_management --> databricks_model_serving["Databricks Model Serving"]
  databricks_model_serving --> legislative_intelligence_applications["Legislative intelligence applications"]
Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component Role Interacts with
Legislative and regulatory data Structured and unstructured source data used for policy and geopolitical intelligence. Data science model development
Data science model development Builds ML and NLP models for legislative outcomes, sentiment, summarization, ETL, and classification. Legislative and regulatory data, MLflow lifecycle management
MLflow lifecycle management Tracks artifacts, builds, notebooks, model versions, and deployment state. Data science model development, Databricks Model Serving
Databricks Model Serving Deploys, hosts, scales, manages, and monitors models without custom serving infrastructure. MLflow lifecycle management, Legislative intelligence applications
Legislative intelligence applications Customer-facing workflows for insights into laws, regulations, sentiment, and voting outcomes. Databricks Model Serving

Data flow

Data scientists build ML/NLP models over structured and unstructured government data, track artifacts, versions, and notebooks in MLflow, then deploy models through Databricks Model Serving. Served models support ETL ingestion, summarization, sentiment analysis, and binary classification, enabling faster updates as legislative data changes.

3 supporting quotes
“MLflow — an open source platform developed by Databricks that helps manage and deploy ML models”article, stated
“reduce the amount of time they spent tracking artifacts, model versions and notebooks”article, stated
“Model Serving has made it significantly easier for FiscalNote’s data teams to deploy a model without needing to worry about building an API, hosting and scaling their infrastructure”article, stated

Why this architecture

FiscalNote chose Databricks because its previous fragmented deployment stack slowed model updates and required extensive custom coding for no-disruption deployments. MLflow and Model Serving gave one lifecycle path for tracking, deploying, scaling, and monitoring models so data scientists could focus on model approach and legislative intelligence rather than platform logistics.

3 supporting quotes
“data teams were hampered by the cumbersome processes required to stitch together various components essential for model deployment”article, stated
“their infrastructure required extensive custom coding efforts to ensure that model updates did not disrupt the operation of existing models.”article, stated
“Now, we can worry less about logistics and focus more on approach.”article, stated

Models used: ML models · NLP models · binary classification models · sentiment analysis models · summarization models

Scale

Alternatives considered (roads not taken)

Manually stitch together model-deployment components

Rejected because: The fragmented process slowed internal workflows and limited model updates to about once per year.

“Before Databricks, FiscalNote had to stitch together various components necessary for their AI model deployments”article, stated
“limited the team’s cadence for deploying and updating certain AI models to around once a year.”article, stated

Custom-coded no-disruption deployments

Rejected because: Custom infrastructure coding increased operational burden and risked data integrity and service continuity.

“This not only increased the operational burden on team members but also posed risks to data integrity and service continuity.”article, stated

Build custom APIs, hosting, and scaling for served models

Rejected because: Databricks Model Serving removed the need for teams to build and scale this infrastructure themselves.

“deploy a model without needing to worry about building an API, hosting and scaling their infrastructure”article, stated

Patterns

observability-tracing (ML lifecycle)

MLflow tracks artifacts, model versions, and notebooks so teams can manage deployment state and reduce manual lifecycle work.

“This helped the data science team reduce the amount of time they spent tracking artifacts, model versions and notebooks to focus on tasks that couldn’t be automated with artificial intelligence.”article, stated

tool-use-orchestration (MLOps platform)

Databricks Model Serving handles API, hosting, scaling, model management, and monitoring for diverse ML workflows.

“Additionally, FiscalNote leveraged Databricks Model Serving, a comprehensive service used for deploying, managing and monitoring machine learning models, whether they are developed by Databricks or sourced from other providers.”article, stated

canary-rollout (deployment safety)

FiscalNote needed no-disruption deployments, and Model Serving reduced custom coding for safely rolling out updates.

“Moreover, FiscalNote’s requirement for no-disruption deployments added another layer to already challenging processes. Simply put, their infrastructure required extensive custom coding efforts to ensure that model updates did not disrupt the operation of existing models.”article, stated

latency-optimization (delivery speed)

The Databricks migration shortened model deployment cadence, letting teams ship three times as many models annually.

“FiscalNote can now deploy 3x the number of models annually, compared to their output before Databricks.”article, stated

Tradeoffs

Standardize on Databricks lifecycle tooling.

Chose MLflow and Databricks Model Serving over Fragmented custom model-deployment infrastructure.

Rationale: The unified platform reduced artifact/version/deployment overhead and let data scientists focus on innovation.

“Thanks to MLflow, FiscalNote teams spent less valuable time and resources on MLOps and invested more in data science.”article, stated
“Now, we can worry less about logistics and focus more on approach.” Because of Databricks tools like MLflow and Model Serving, FiscalNote reduced bottlenecks with their data platform teams to support innovation at scale.”article, stated

Use managed model serving.

Chose Databricks Model Serving over Custom APIs, hosting, scaling, and no-disruption deployment code.

Rationale: Managed serving improved deployment speed and reduced the operational burden and continuity risks of custom infrastructure.

“without needing to worry about building an API, hosting and scaling their infrastructure or planning no-disruption deployments.”article, stated

Gotchas (what bit them)

Fragmented MLOps creates yearly deployment cadence.

Root cause: Teams had to stitch together artifacts, builds, versions, APIs, hosting, scaling, and custom deployment code.

Resolution: Adopt MLflow and Model Serving to unify lifecycle tracking and model deployment.

“cumbersome processes required to stitch together various components essential for model deployment, such as tracking artifacts and builds.”article, stated
“Using this feature alone, FiscalNote experienced faster time to market and scaled up their model deployments to an as-needed basis.”article, stated

No-disruption deployment requirements can increase operational risk.

Root cause: Custom coding was needed to prevent model updates from disrupting existing models.

Resolution: Use Model Serving to deploy models without custom API, hosting, scaling, and no-disruption infrastructure.

“their infrastructure required extensive custom coding efforts to ensure that model updates did not disrupt the operation of existing models.”article, stated
“Model Serving has made it significantly easier for FiscalNote’s data teams to deploy a model without needing to worry about building an API”article, stated

Poor data-asset discoverability slows model development.

Root cause: Data scientists could not easily access necessary data assets.

Resolution: Use a unified data analytics platform so teams spend less time on platform logistics and more on model approach.

“a lack of discoverability of essential data assets made it difficult for data scientists to access and use the necessary data”article, stated

Evaluation

FiscalNote measured deployment productivity and time-to-market improvements after moving model lifecycle and serving work to Databricks.

Metrics: Model deployment output: 3x annually compared with before Databricks. · Previous update cadence: around once per year. · Previous deployment phases: several weeks.

“FiscalNote can now deploy 3x the number of models annually, compared to their output before Databricks.”article, stated
“Databricks allows our team to build and deploy ML models faster, reducing time to market and allowing our employees to be more productive”article, stated

Outcomes

FiscalNote serves nonprofits, government agencies, and over half of the Fortune 100; the Databricks shift enables three times as many annual model deployments.

“By adopting Databricks, the business streamlined model deployment, making the process 3x faster and increasing analyst productivity.”article, stated
“This newfound agility enables FiscalNote to process and analyze legislative data quickly and accurately. As a result, they can deliver timely, actionable intelligence that helps clients swiftly adjust to new laws and regulations.”article, stated