Quality assurance · Production

Crisis Text Line uses Databricks to centralize data and deploy LLM-powered crisis counselor training

The problem

Crisis Text Line had a fragmented, siloed data infrastructure where business rules and context were lost across silos, preventing downstream teams from making efficient real-time decisions as the organization scaled to support over 1.3 million conversations per year.

First attempt

Legacy infrastructure suffered from query timeouts, batch-script failure points, and siloed data that led to duplicated efforts and an inability to build a cohesive data culture across the organization.

Workflow diagram · grounded in source

Data centralization to single source

integration

“enabling Crisis Text Line to centralize its data and create a single source of truth”

Granular access governance

integration

“Unity Catalog has been instrumental in managing data access at granular levels, ensuring compliance and enabling faster data-driven decisions”

Model lifecycle management

ai_action

“MLflow is used for managing the full model lifecycle, from training to deployment. It tracks experimentation and model versioning, ensuring that the right models are released in the correct capacities across workspaces.”

LLM counselor training simulator

ai_action

“A large language model (LLM)–powered conversation simulator allows new crisis counselors to practice engagement strategies in difficult scenarios, boosting their confidence without real-world repercussions.”

Conversation phase classification

validation

“a conversation phase classifier under development will help the team level up conversation assessment to maintain high quality and ensure appropriate responses”

Program health dashboard

output

“using a suite of data product dashboards to monitor program health and quality”

Reported outcome

Databricks enabled Crisis Text Line to centralize data into a single source of truth, dramatically reduce dataset availability time for clinical teams, deploy an LLM-powered counselor training simulator used by over 50 volunteers and clinicians, and reduce infrastructure overhead while improving program ROI.

Reported metrics

Conversations supported in past yearover 1.3 million

Volunteers and clinicians using simulatorover 50

Time to make critical datasets availabledramatically reduced

Infrastructure and engineering overheadreducing

Show all 6 reported metrics

conversations supported in past yearover 1.3 million

volunteers and clinicians using simulatorover 50

time to make critical datasets availabledramatically reduced

infrastructure and engineering overheadreducing

program ROIincreased

research and impact workaccelerated

Reported stack

DatabricksUnity CatalogMLflowSpark Declarative PipelinesLlama 2

Source

https://www.databricks.com/customers/crisis-text-line

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Databricks enabled Crisis Text Line to centralize data into a single source of truth, dramatically reduce dataset availability time for clinical teams, deploy an LLM-powered counselor training simulator used by over 5…

What tools did this team use?

Databricks, Unity Catalog, MLflow, Spark Declarative Pipelines, Llama 2.

What results were reported?

Conversations supported in past year: over 1.3 million; Volunteers and clinicians using simulator: over 50; Time to make critical datasets available: dramatically reduced; Infrastructure and engineering overhead: reducing (source-reported, not independently verified).

What failed first in this deployment?

How is this quality assurance AI workflow structured?

Data centralization to single source → Granular access governance → Model lifecycle management → LLM counselor training simulator → Conversation phase classification → Program health dashboard.