Scaling Human Connection with Real-Time AI Conversation Monitoring

Summary

Crisis Text Line built a Conversation Support Engine on the Databricks Data and AI platform—an ensemble of AI models that provide real-time metadata extraction from crisis conversations, tagging clinical risk signals, safety dimensions, and policy compliance patterns to strengthen supervisor visibility.
The system migrated from fine-tuned models to retrieval-augmented generation using Databricks vector search and foundational LLMs, improving the accuracy and maintainability of safety assessment and risk classification models.
A simulation-based volunteer training platform powered by AI-driven conversations with real-time feedback reduces volunteer drop-off and builds responder confidence, extending the nonprofit's capacity to support people in crisis.

Scaling Human Connection with Real-Time AI Conversation Monitoring

Real-time conversation monitoring is critical to crisis intervention at scale, but manual supervision limits response speed and creates visibility gaps. Crisis Text Line built a Conversation Support Engine on Databricks, an ensemble of AI models providing real-time metadata extraction from conversations. These models tag clinical risk signals, assess safety dimensions, and surface policy compliance patterns. The system strengthens human judgment rather than automating it, giving supervisors and volunteers precise visibility to focus on what matters most.
this video covers the architecture: a two-layer approach using message-level metadata aggregated into business logic. You'll learn how Crisis Text Line migrated from fine-tuned models to retrieval-augmented generation, leveraging Databricks' vector search and foundational LLMs. The talk includes their deployment strategy with gradual rollouts and training wheels, and demonstrates their engaging training platform, a simulation-based system using AI-driven conversations with real-time feedback to build volunteer confidence and reduce costly drop-offs.
🤝

Chapters

FAQs

What is the Conversation Support Engine at Crisis Text Line?

The Conversation Support Engine is an ensemble of AI models that runs in real time alongside crisis conversations, extracting metadata such as clinical risk signals, safety dimensions, and policy compliance patterns. It gives supervisors and volunteer responders precise visibility into which conversations need attention, without replacing the human-to-human connection at the core of the service.

How does Crisis Text Line use AI without replacing human volunteers?

Crisis Text Line uses AI strictly to strengthen the human element rather than automate it, because their research shows that people in distress want a real human at the other end. The AI models surface information to human supervisors, allowing them to focus their attention more effectively across many simultaneous conversations.

Why did Crisis Text Line migrate from fine-tuned models to RAG?

The migration to retrieval-augmented generation using Databricks vector search and foundational LLMs improved the ability to handle the variety of language patterns encountered in crisis conversations without requiring constant fine-tuning cycles. RAG approaches also reduce maintenance burden compared to keeping fine-tuned models updated as language and context evolve.

How does the simulation-based training platform help Crisis Text Line volunteers?

The simulation platform uses AI-driven conversations to provide volunteers with realistic practice scenarios and real-time feedback during training, helping them build confidence before supporting real texters. Reducing volunteer drop-off during training is critical because trained volunteers also improve mental health outcomes in their own communities, with over 95% of crisis responders reporting that volunteering improved their own mental health and resilience.

Full transcript

[00:09] Hello everyone and thanks for coming and joining this session and thanks for Databricks for the invitation. I'm Mateo from Crisis Text Line. Crisis Text Line is a global nonprofit uh organization that offers free 24/7 mental health crisis support, okay?
[00:24] And today I want to walk you through how we use AI as a tool to strengthen human-to-human connection and how a small team is able to build these models and ship them within the Databricks ecosystem.
[00:41] Oops. Okay, yeah, so I'm Mateo Garcia Pepino. I'm the lead AI engineer at Crisis Text Line and of course I'm very much involved in all of these efforts to put AI on our platform and on our training experiences.
[01:01] Let me start with some slides on Crisis Text Line, our mission, our service and ultimately also why do we need AI to scale our operations, okay? So Crisis Text Line was founded nearly 13 years ago to address the lack of text-based text-based mental health
[01:17] crisis support, okay? At the time where 73% of teens in the US had a smartphone and were sending roughly 100 text messages a day. We offer free 24/7 confidential support to anyone, anywhere that is
[01:33] that navigating a moment of distress, okay? And we use technology to help people in distress reach another human rapidly, efficiently, and effectively. Okay. So, texters in crisis are supported on our bespoke platform by
[01:50] trained volunteer responders under the live supervision of trained clinical staff, okay? And there is one thing that matters a lot to us is that we don't use technology and specially AI to replace humans, okay? We use it we use it to strengthen the human to human element.
[02:07] And this matters to us a lot because one of the most common questions that we get on our platform from people that reach out is whether they're talking to a bot or a real person. What we see is that people want a real person at the other end of the line that genuinely cares.
[02:24] Okay? This is very important to us. And this also got both got both ways. Our research shows that the volunteer experience is powerful on its own. Over 95% of our crisis responders say that their own mental health and resilience improve through volunteering and that
[02:40] they've gained skills to support someone in their families, schools, workplaces, and communities. Over 90% have done so in the past 6 months. 42% do so regularly and we conservatively estimate that for every conversation on our
[02:56] platform a volunteer supports or prevents another crisis moment in everyday life. To put everything in context, brain disorders and mental ill health cost the global economy $5 trillion a year and this number is projected to reach 16
[03:12] trillion in the next 5 years, okay? So, let me give you some numbers on the impact that we've had so far. Particularly for the US, we've served texters since 2013 for over 12 million
[03:27] conversations with people in crisis across all 50 states plus DC and Puerto Rico. 69% of our texters are 25 and under, 44 identify as people of color, 46 as LGBTQ+, and over 70,000 conversations
[03:43] have been handled by our Spanish language service. Now, let me go to our international impact. We are expanding this model worldwide, right? So far, our technology and model have supported over 17 million conversations globally across 24
[04:00] countries and three languages through partners like Shout in the UK or Aquí Estoy across Latin America and Spain. To date, this model of free confidential 24/7 support has been accessible to about 12% of the world's population.
[04:20] Okay. So, a bit more context on exactly what is the service that we provide, right? How does the service actually work? Basically, anyone, any age, in any kind of crisis can reach out by SMS, WhatsApp, or website.
[04:36] We triage conversations at the outset to move the most acute cases to the front of the queue. A trained volunteer crisis counselor then collaborates with the texter to de-escalate and find healthy next steps. Throughout, a trained clinical supervisor monitors the conversation in
[04:52] real time, giving feedback, and can step in if additional support is necessary. Now, just a quick slide on the technology that powers all of this, right? This is a high-level picture of how our crisis tech tech care technology
[05:08] uh brings together the three things uh that I was referring to earlier. First of all, our crisis conversation platform, then our volunteer ecosystem, and of course the data ecosystem. These are the main three components of the technology that we built, and everything that we do
[05:24] with with AI is built on top of these these three things, okay? Now, after this quick introduction, let me go and present to you what is the problem that we have and what is why is why we need to leverage
[05:39] AI, okay? Basically, the problem is that crisis intervention at scale is very is very challenging, okay? One of one of the examples of this is uh
[05:55] conversation monitoring. Basically, monitoring the state of conversation, actually seeing what's happening in real time across live conversations, is critical to provide the best possible service. And today, this is a fully manual
[06:10] exercise. Basically, a supervisor has to watch many conversations at once and track risk throughout each one. They're manually distinguishing high-risk from low-risk moments, and crucially, some moments where risk goes up or down in the conversation.
[06:26] Because everything is manual, teachable moments get missed, and volunteers who need support don't always get it when they need it, okay? And here, our bet is pretty simple. If we surface more information through AI, supervisors and volunteers would be
[06:44] able to spend their attention on what actually matters, which is the connection with the texter. So, this is where we think AI is a lever for us, okay? Again, a lever, not a replacement.
[07:00] I think we think that it's a lever to save more lives, as I'm writing there, to train better volunteers, and to ultimately extend our impact globally. Why do we think that we're uniquely positioned to do this? Because of the these three things that you see on the screen, right? First, we have the data.
[07:17] We have 13 plus years of signal-rich crisis conversation data for more than 300 million messages exchanged. And ultimately, this means that we can teach models what crisis conversation looks like because we've actually had millions of them on the platform, okay?
[07:34] Second, we also have the clinical expertise. We build these models together with clinicians as multidisciplinary efforts, basically. So, we have trained clinicians that are helping us make sense of the data and align the models to the reality of
[07:49] crisis intervention practice and clinical value. And third, we are also running an active service, meaning that we have a live service to deploy, test, and harden these models against. This is pretty beneficial for us to build the necessary feedback loops with
[08:05] supervisors and volunteers who are actively providing care and do the necessary gradual rollouts before anything touches highly consequential workflows, okay? As you can imagine, the service that we provide is highly consequential, is high stakes. So, we cannot put everything in
[08:22] production right away, and we need to think very deeply about the consequences of our deployments and our AI implementations. Okay. So, now I will introduce what we call the conversation support engine, which is basically the incarnation of
[08:37] this idea of utilizing AI as a layer that provides real-time monitoring and real-time visibility into what is happening in conversations. And basically, the conversation monitoring engine is is just at its
[08:54] core, it's an an ensemble of models, right? Each of them extracts clinically relevant features from conversations. Some features can be extracted at the message level, it's message metadata, basically. Some of them can be extracted at the conversation level. For instance,
[09:11] the summary of a conversation so far. On the left of the of the the screen, you can see the shape of it. We basically take a conversation and its messages, and we build a metadata layer on top of it, okay?
[09:27] Our models are tagging messages for different clinically relevant features. For instance, the risk level of the conversation so far, uh tags on safety assessment dimensions, which is something that I will introduce later on. It's a policy that the volunteers follow as they have conversations with
[09:43] texters, or even tax for what we call good contact techniques, which are basically linguistic figures that are utilized throughout the conversations to show texters that we're actively listening to them and that we're showing empathy to their crisis, okay?
[10:00] The idea here is that each model produces a structured output that we can aggregate, query, and act on depending on the use case. And the important thing here is that the output of these models is low level enough that we can power many different products out of them, okay?
[10:17] Also, as a byproduct of all of this and putting these models in production as part of workflows is that every time we release a new feature, we are basically creating and enriching a layer of metadata over our crisis conversations that then we can use for analytics and research purposes.
[10:39] Okay, so what I just said, I think it's quite important, which is basically that at the core of our model building philosophy, we have this idea of building one foundation that is able to enable many different products, okay? So, we treat these low level features as a shared asset and not a per product cost. So, we are not building a model
[10:56] for a single feature, but rather we are building a model that makes clinical sense or sense at the policy level, and then we can turn it into different products, okay? So, for instance, take these safety assessment dimensions. Basically, the safety assessment policy
[11:12] is the policy that volunteers go through to assess the risk of suicide of a texter that reached out, okay? This policy has many different components, and basically the model that we've built is able to tag messages for
[11:27] whether they contain information on these different components, right? So, the idea is that by tagging for these different components, we would be able to answer different questions at different levels. For instance, in a training setting, we would be able to
[11:43] tell a volunteer whether they are correctly applying the safety assessment money process procedure. Or if we were on the platform, we would be able to flag for a supervisor uh a conversation from a counselor that is not going properly through the some
[11:58] procedure, right? So, it's one foundation and then it can enable different product uh incarnations, okay?
[12:13] Oops, I Now, just a few slides on why Databricks is important for us and why we're building this on the Databricks stack. Um which is basically this very, very high-level diagram. Uh please bear with me. It's just to show how a small team
[12:28] like ours is able to operate a platform like this without creating dependencies uh with other teams, right? Generally, we lean on the general ML Ops stack that Databricks has to offer, where Unity Catalog and the Lakehouse are our single source of truth for de-identified
[12:44] conversations, evaluation data, and model artifacts. Then, we use MLflow heavily to log everything, every experiment, every evaluation, every run, and also the traces. Okay? And that really gives us a consolidated source of truth for
[13:00] evaluation and iteration of models. Also, one thing that we're doing is that now we're shifting our model building and deployment approach from like a fine-tuned model towards a more paper token leveraging foundational APIs. So, we are also using all of the model
[13:15] serving functionality, vector search, and all of that to deploy these uh new architectures that we're leveraging, okay? Of course, there are a few things that do happen outside of Databricks. We do have AWS Lambdas that are running on top of the LLMs that we serve in Databricks and that connect
[13:32] sort of like that interact with the UI. But I think that the takeaway here is that most of like almost everything is is running within the Databricks ecosystem, okay?
[13:54] really benefiting from this serverless Databricks functionality is this migration that we're doing from provision throughput sort of like fine-tuned models to rag-like approaches. Basically some of our models do work out of the box, right? LLMs are very powerful and
[14:11] they are powerful enough to understand the nuanced clinical policy that we have on our that we run on our platform with good prompt engineering and solid few-shot examples, right? But it is true that some of the clinical
[14:27] policies that we follow are nuanced enough that the model doesn't get there alone, okay? So for those we then need to provide more context for the model via fine-tuning or standing up a rag-like architecture, right? So that's really where we move
[14:44] from reasonable demo-like accuracy to something that is product-ready that we can put in front of users, okay? And one thing that we learned through the past year is that for deployment and scalability purposes running a rag agent is more convenient
[15:02] than having to maintain and deploy fine-tuned models. So we're in the middle of this migration, we like this message tagging model that I keep referring to is now a fine-tuned Llama 3.1 8 billion parameter model
[15:17] that is deployed via provision throughput and now we're moving it to a rag uh um a rag approach that utilizes vector search and Claude Sonnet like that's the best candidate so far for uh re-ranking like algorithm. So, um just to show
[15:34] how we're benefiting from from this minute's uh feature set. Okay. Now, let me uh spend the rest of my time talking about one application of this
[15:49] conversation support engine, right? Which is basically what we are calling so far engaging training and is really the idea of building more compelling practice scenarios for volunteers to tackle the real challenge of managing and scaling uh
[16:04] a volunteer community. So, going back to crisis intervention at scale, right? Crisis intervention at scale is quite challenging. And one of the most challenging aspects of
[16:21] of of it is basically maintaining a community of building and maintaining a community of trained volunteers, right? We do know that volunteer power models are the way to deliver efficient, high-quality crisis intervention at
[16:37] scale. That we know. But the challenge of building that volunteer community and maintaining it is pretty hard. So, as we grow, we need to improve the efficiency while holding the quality bar. Okay? And
[16:52] basically what we require is best-in-class training and practice systems that build skills and confidence across every phase of the volunteer journey journey, okay? From early training to once they're on the platform talking to real to real texters and they need to be
[17:09] coached. Okay? So, let me show you some numbers on why this matters so much. Um these are our drop-off rates across the volunteer journey. Basically, out of 100 volunteers that
[17:25] start our training, about 50% 50 graduate, 35 take a shift, 29 take multi multiple shifts, and only 21 of them is continues to volunteer after 3 months. Okay?
[17:40] So, that's roughly an 80% drop-off from the start of the training to the 3-month mark. Of course, we don't want this. We want to enable volunteers to be the most impactful volunteers that they can be. Okay?
[17:59] So, we have this hypothesis around these engaging training exercises that could uh help us work through this challenge. Basically, our current curriculum is very strong at giving a foundation in core concepts to volunteers and core policies uh that need to be followed on
[18:16] the platform, but it doesn't consistently give volunteers with a chance to practice, right? And that gap in practice is ultimately becoming a gap in confidence. And what what I wrote there is something that we we are being reported consistently. Volunteers are reporting
[18:33] to us that they feel too nervous, that they're not ready, that they're not ready to engage on the platform with real texters, right? So, how do we tackle that? Basically, the bet here is that scaffolded individualized simulation-based exercises
[18:49] can help us build competence and confidence on our volunteers. Okay? So, let me use a an analogy for unit like a unit economics analogy, which is basically
[19:05] that what we want to do is increase the volunteer lifetime value. Okay? We want to help everyone who starts with us become the most impactful volunteer that they can be. Okay? Of course, we're a nonprofit, so unit
[19:20] economics doesn't make the sense that it makes in for-profit businesses, but I hope I hope that this analogy gives you a perspective on what are we trying to do. Okay? Also, by using AI to build these simulations and deliver feedback, um one
[19:37] thing that we can also do is that reduce the burden on our coaching and training teams so that we can ultimately also lower the acquisition cost. Meaning that we can maximize the LTV to CAC ratio to have the most efficient and effective uh
[19:53] volunteer funnel that we can. If we maximize the LTV to CAC ratio, then we are decreasing the cost per conversation, and then we are unlocking uh global scale without sacrificing quality, which is the bar that we need to always keep in mind.
[20:11] So, this is basically what we're trying to do with these engaging exercises. So, more concretely, uh what we're doing is that we're building these simulation-based exercises, and they can have different flavors. I will now present one of them, but the core is the same. Trainees
[20:26] practice conversations with an AI-driven texter that uh with an AI-driven texter, sorry, and then as they're having the conversation, they get clinically grounded, policy-aware, real-time feedback that is calibrated to our crisis interventions practices.
[20:43] Okay? And the idea is that we can also embed this in our learning management system, so it's a core part of the training journey and not a side tool. Okay? Of course, the more volunteers practice with this type of tooling, the more data that we gather so that we can help our
[21:00] uh team of training specialists and coaches improve the training experience. Um I have a few more slides Uh uh before I go into the demo. Um this one I I wanted to show because I
[21:15] think that while what I'm showing here is really nothing um new, I think it's it's worth saying it out loud. It's basically like what I what is our strategy or what is our approach to deploy this safely um
[21:33] on the platform without having unintended consequences, okay? And again, I think that this deployment strategy or this deployment approach is something that holds not only for training and what I'm uh talking about now, but also for
[21:48] platform deployments, even more for platform deployments, right? Basically you we need to think about gradual rollouts, of course, because models can fail and will fail. So, you need to start with training wheels and user augmentation and not use user
[22:05] automation, and then move slowly and steadily towards more automated assessments, okay? And then, of course you need to learn from your implementations. And that means that you need to keep high throughput and a high iteration speed, so you learn from what
[22:21] you deploy and you're able to harden not only your models, but your model implementations and your product implementations, okay?
[22:37] Um okay, so this is generally how the inference process works for these engaging exercises. Uh and I wanted it I wanted to show it because I think it's pretty general and also gives a feel for how could how this could work not only for training, but also on the platform
[22:52] uh with real conversations, okay? It's very transferable to whatever uh can be deployed on the platform. It basically for this particular case of training, it starts with our simulation scenarios, which I actually presented here last year. Um these scenarios are built to mimic real
[23:10] texters with real issues to make practice much more immersive, okay? And right now, all of them are running in Cloud Sonar 4.5 through a prompting engine Um then what we're doing is that we're
[23:26] monitoring messages. And again, on the platform, this would basically like the shape of the monitoring would be the same, but on top of real conversations and not simulations. The schema is the same, basically. So, as the conversation unfolds, the models are listening in and tagging messages
[23:43] for relevant features as they go, okay? Here, I'm just mentioning two of them. Uh the first one is the one that I will demo, but uh the first one is basically this model that I keep referring to, which is
[23:59] this safety assessment model that tags uh tags messages for this uh safety assessment model dimensions. And the second one is like this good contact techniques uh model, which tags messages for whether good good contact techniques show up on the message or not.
[24:15] Okay? You have your message level stream with all of these tagged messages. And on top of that, what you have is at the conversation level, we are building what we call the business entities, right? So, we have the message level
[24:31] metadata stream, and then we build on top of that the business entities. The the things that make business sense. So, basically, the message level features are deliberately low level. They make sense at the clinical level, right? And it is this second this second layer that
[24:46] applies the logic that makes sense at the business level, okay? For instance, in the case of disengaging exercises, it decides whether some criteria that needs to be met for the exercise is met or not, okay? So, that means that if certain tags appear on the message level
[25:03] stream, the criteria gets marked as complete or not depending depending on the practice on the practice setting and the the particular message tags that we're looking at, okay?
[25:19] So, basically this two-layer approach is interesting because we the message level stream is our metadata layer and on top of that we're applying the business logic that can enable different things. Okay, now let me finally give you more context on what is the safety assessment
[25:35] model, okay? Before we go into the demo so that you know why it's so important to teach volunteers about these policies and why it's important to give them opportunities to practice this, okay? So, basically the safety assessment
[25:51] model is essentially the clinical policy that volunteers follow. It guides them through assessing risk for any texter who might who might be in crisis, okay? And the policy has two main parts to it, okay? First, the volunteers need to ask the
[26:06] safety prompts. There are two mandatory questions that need to be asked early in the conversation. These two questions are you can see them there are explicit questions about whether the texter is having thoughts of suicide or whether the texter um
[26:22] intends to engage in self-harm, okay? If by if after asking these questions, there are any positive responses, if there is any positive response from the texter, basically the volunteer needs to
[26:38] Sorry. The volunteer needs to continue working with the texter to explore the four dimensions that you see on the right. Those are the four dimensions associated with the safety assessment model and are the four dimensions associated to the suicidal ideation of our texter. Desire, intent, capability, and buffers.
[26:55] Okay? So, we've built a model basically that is able to tag messages containing these safety prompts, as well as messages that contain information or signal on these different four dimensions. Okay? And again, I cannot stress enough how
[27:12] following this policy correctly is critical, right? This is how we provide the best possible service to texters and keep them safe. Okay? Keep them as safe as possible. So, basically, this is why training volunteers and giving them opportunities to practice is so important for us.
[27:31] Okay. So, now I will go into the demo, finally, which is basically an exercise that teaches um volunteers how to go through asking the safety prompts properly. Just that.
[27:55] So, this is the exercise. I have like the learning objectives there. I have context on the scenario. And then I have uh words on the fact that this is utilizing AI, of course. Um on the left side of the conversation window, you can see the criteria that needs to be met. And on the right side, you will see that real-time feedback
[28:12] appears and updates depending on how uh the volunteer is going through the conversation. Okay? So, there are four criteria that uh volunteer needs to follow to be asking the safety prompts correctly, right? The two questions must must be asked with
[28:27] the correct wording. There cannot be deviations to the wording that I was showing before. The timing must be right. It cannot be too early because rapport with the texter needs to be built first. And it cannot be too late, either. And also, finally, the two questions
[28:44] must be asked close together. Basically, they cannot be asking the same question, but they cannot be asked five or six messages apart. They need to be close together, okay? So, what is going to happen here is that
[29:00] without building enough rapport, I'm going to ask the first safety prompt, the one that ask about thoughts of suicide. I'm going to ask it verbatim, right? No deviations to the wording, but I'm going to ask it too early. So, that means that the criteria on wording it's going to be met, but the criteria on
[29:16] timing it's going to be not met. And I will get also the feedback saying that I did rush to ask the safety prompts, and I should sort of like next time uh make sure that I that I do that um a a bit later in the conversation, okay?
[29:36] There is a delay, a synthetic delay that we introduced to make it feel more uh realistic because of course the model has already responded. You can see the criteria have met and not met. And then it is basically telling me that
[29:52] I failed the timing requirement also as part of the feedback uh column. And now I'm going to ask the second the second safety prompt, but I'm going to ask to ask it with a slight wording change.
[30:07] The The second safety prompt basically ask about self-harm today. And I'm going to write self-harm lately, okay? And what is basically going to happen is that the the criteria on
[30:22] writing the safety prompt verbatim is not going to be met, while the criteria of asking the two safety prompts close together is going to be met, okay? And I will also get a feedback blur that tells me why that is not correct and why I should
[30:38] uh try to uh do it better next time. Then I have a summary and this is basically an exercise on asking safety prompts, right? I have the summary, what is the criteria I met, what is the criteria that I haven't met.
[30:53] This is basically the idea across the whole training rubric. We can teach these very specific policy pieces through these type of exercises and we can make it such that at some point is not even an exercise on something concrete, but rather
[31:10] a whole simulation with training wheels on top where I'm getting feedback along the way. And you can imagine how this can also be transferable to a platform setting what we're enabling what we are enabling our co-pilot-like
[31:25] experiences that really help the volunteer in the real setting. But of course the fact that we're first teaching is not only good for like training purposes, but also it's good for us to learn how uh users interact with these types of
[31:41] features and how we could enable them on the platform with real conversations in a high-stakes setting. And that's basically it. Um let me close where I started. Uh
[31:56] At Crisis Text Line we deeply believe that AI is really the the strongest lever that we have to scale human-to-human connection. And that's the work that we do and that's the work that we we tend to keep doing. And I have nothing else to say. Thank you so
[32:13] much again for for coming here. I If you have any questions, I'm here to answer them. Um if you're interested in uh volunteering with us, uh please do not hesitate. It's like a very impactful community service, so reach
[32:29] out and there are some links if you want to reach out. Uh thank you thank you again.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.