Thumbtack Powering Safe, Smart Home Services on Databricks with GenAI

Building the Most Trusted Home Care Platform

Thumbtack’s mission is simple but ambitious: empower people to manage their homes confidently and effortlessly by making every service, repair, and improvement reliable and safe. We support local economies by connecting millions of homeowners nationwide to over 300,000 skilled professionals, from plumbers and electricians to wellness providers and event organizers. The opportunity is vast, but so is the complexity — our goal is to guarantee consistent, exceptional results for every customer, every time.

Unlocking GenAI Value at Thumbtack

The rapid evolution of home services and rising customer expectations mean we are continually advancing our platform — data volumes, unpredictable customer and professional needs, and expanding service categories present technical and organizational challenges. Thumbtack faced fragmented data science and engineering workflows, siloed infrastructure, and a high bar for privacy and safety.

Solving these challenges required more than clever algorithms or faster infrastructure. It required a connected, trustworthy data and machine learning platform that puts safety, privacy, and collaboration at the core. Our approach: unify our GenAI ecosystem on top of Databricks to drive real, measurable impact.

Trusted GenAI, Centralized Security, and Productive Data Science

Elevating Trust and Safety with Fine-Tuned LLMs

Thumbtack’s semi-automated message review pipeline is the backbone of our digital trust platform. Each message, between a customer and a pro, is screened by both a rule-based engine and a machine learning model. While typical abuse cases can be caught with simple rules, many nuanced policy violations cannot. Early systems based on Convolutional Neural Networks (CNNs) struggled to differentiate between sarcasm, context, or implied threats.

Fine-tuning large language models on Thumbtack’s own labeled data made a step-change difference. With our hybrid workflow, a CNN model pre-filters for obviously good messages, reducing LLM workload by 80%. The fine-tuned LLM then focuses its power on the most challenging 20%, increasing detection precision by 3.7 times and recall by 1.5 times. Tens of millions of messages are processed each year, ensuring conversations remain safe while maintaining honest interactions and avoiding unnecessary costs.

Building on Databricks: Secure, Standardized, and Flexible

All advanced AI and trust workflows at Thumbtack now run through a unified ML platform built on Databricks. Key investments and safeguards include:

• Centralized LLM workload management: By running all GenAI workloads on Databricks, we reduce our attack surface and maintain a consistent governance model.
• Workspace isolation: Virtual private clouds ensure sensitive data stays protected, with granular permissions managed through tools like Terraform. We use Unity Catalog to enable serverless and Databricks Genie to access BigQuery, as part of how we ensure safe permissions management.
• Automated privacy protection: Open-source and internally developed scrubbers remove Personally Identifiable Information (PII) and confidential information from data as it flows through notebooks, models, and pipelines.
• Comprehensive observability and monitoring: Every model, notebook, and API route is tracked for data drift and PII exposure. Visualization tools confirm that risky data is not leaking into downstream systems.
• Centralized secrets and artifact management: With MLflow and secrets managers, teams manage credentials securely, version all models, and collaborate productively — no more decentralized, brittle copy-pasting of keys or libraries.

Best Practices in GenAI Operations

• Hybrid AI workloads: Production services run on AWS with analytics on Google Cloud, but all GenAI workflows are centralized and standardized for reproducibility.
• Reuse and efficiency: MLflow and notebook tracking mean experiments or solutions can be shared, compared, and extended across engineering, SRE, and analytics — all with consistent privacy controls.
• Proactive privacy safeguards: Thumbtack customizes open source PII scrubbers to its specific needs and enforces monitoring at every layer. Industry trends indicate that PII-related notebook and model breaches have increased by 300% since 2022, making these protections business-critical.

More Safety, More Trust, More Innovation

• Marketplace scale: Millions of U.S. users and 300,000+ local service businesses now interact within a platform that prioritizes security and reliability.
• Superior message filtering: Precision up 3.7x, recall up 1.5x, costs controlled by processing only the riskiest 20% of messages with LLMs while safeguarding privacy at every step.
• Collaboration and efficiency: Centralized, reproducible ML workflows eliminate manual handoffs and enable rapid cross-team innovation, allowing data scientists, SREs, and ML engineers to work in sync.
• Confidence in scale: With robust technical and process controls, Thumbtack delivers on its mission to be the most trusted, transparent marketplace for home services.

As Thumbtack continues its GenAI journey, every team is empowered to experiment, collaborate, and deliver safer, smarter home service experiences. The strategy is grounded in real-world impact, demonstrating how AI, privacy, and platform thinking combine to create value for both professionals and homeowners.

Watch the Thumbtack Boosting Data Science and AI Productivity With Databricks Notebooks 2025 Data + AI Summit presentation.