Taki Hasegawa, The Trevor Project

Thank you for tuning in to today’s episode of the Breaking Changes podcast. I’m your host and Chief Evangelist for Postman, Kin Lane. With Breaking Changes we explore specific topics from the world of APIs, but through the lens of business and engineering leadership. Joining me today we have Taki Hasegawa, senior machine learning engineer at the Trevor Project. Taki expanded on the stories I have regarding how APIs and machine learning can make a difference in our world, and I was captivated by how they are using AI for training up the next generation of LGBTQ counselors. Here we are, let’s dive in.

Start with the basics. Who are you, and what do you do?

Yeah, so my name is Taki Hasegawa, and my pronouns are she/they. Personally, I love PC gaming, meditating, and I have a cat named Simon. And professionally, I’m a senior machine learning engineer at the Trevor Project. And just for some background, the Trevor Project is the world’s largest suicide prevention and mental health organization for LGBTQ young people.

So very important. Thank you for being here today. So I would like to understand more about the role of technology at the Trevor Project and the impact that it makes there. Can you share a little bit more?

Yeah, for sure. So the role of technology at the Trevor Project is to help both transform our life-saving crisis services, which we offer three of, TrevorLifeline, TrevorText, and TrevorChat. And just a bit more background, these crisis services, LGBTQ youth who are experiencing a crisis can reach out to volunteers through these services, so they can either call us, they can text us on their phone, or they can chat with us in a browser. So again, the role of technology at Trevor is to transform these crisis services and to also increase its capacity to serve more youth every year. More concretely, what this looks like is, we’ve expanded our digital crisis services to be 24/7. It wasn’t always that way, and there was a significant investment in technology and elsewhere to do that. And more recently we’ve been using AI to do a variety of things, including using AI to connect youth with the highest risk of suicide to counselors as quickly as possible, to support the moderation of our social networking site to help keep youth safe, and, as we’ll talk more today, we’ve also created an AI chatbot called the Crisis Contact Simulator. And this chatbot is used to help scale our onboarding and training of thousands of volunteer crisis counselors per year. So we’ve been pretty expansive in how we use technology at the Trevor Project, and we’ve been pretty proud of what we’ve done.

Yeah, definitely a very powerful use of technology. I see many different industries, so it makes me happy to see it used for such an important area of our culture and society. I’m really interested in the Crisis Contact Simulator. What was the need for this, for helping train and bring people up to speed so that they could help out?

Yeah, exactly. So the Crisis Contact Simulator is an AI chatbot that we developed to provide simulated role plays that aspiring counselors can use to practice the skills they learned. So prior to CCS, or the Crisis Contact Simulator, all of our practice conversations for these aspiring counselors were led by instructors who would act as LGBTQ youth. So what we found was a large investment on our part of where we concentrate our resources. Conducting these role plays was taking a lot of time for our teams. So while we still have some instructor-led role plays, we were able to replace two of them with the Crisis Contact Simulator. So this had a multitude of benefits, including helping scale our training services. I think before the Crisis Contact Simulator we trained roughly 30 new counselors each quarter, and now with the Crisis Contact Simulator and other initiatives that we’ve implemented at Trevor, we now trained a thousand digital volunteer crisis counselors last year. And that increase was a result of a lot of projects, of which the Crisis Contact Simulator was a key part. So that’s the role that CCS played for that.

And we actually went through a pretty extensive development process with Google.org to build this AI chatbot. So Google.org, which is Google’s philanthropic arm, provided us with some funds and Google fellows to help us build up the CCS. And with them we tried a bunch of different AI models. We ended up with OpenAI’s GPT-2, and we were able to create realistic conversations in which the CCS acts as the LGBTQ youth. And through these realistic conversations, like I said, yearly a thousand digital volunteer counselors practice the skills they need before they talk to actual LGBTQ youth.

Wow, what a positive use of AI. I would say a lot of, well, the majority of AI use cases I come across, some of them I don’t feel are very positive. So did you know much about AI before this? What’s your general view of artificial intelligence? Were you skeptical of how it could be applied before this project, or are you feeling more optimistic about it now?

Yeah, thanks for asking this. So I’ve been involved personally in data science and machine learning and AI for a long time now. I would say for me, CCS was definitely the project that had the most social impact. And again, I’m really proud, not only proud, I’m really grateful to have been part of a team that has been able to work on such a project. With that being said, my view on AI right now is that it’s a tool, a powerful tool that helps empower its users. With that being said, nowadays I really try to prioritize and I applaud use cases of AI for social good, because I think there’s a lot of benefit to be gained from using AI in that manner.

Yeah, it makes me happy to see it used for social good, because I think it gets used behind the scenes and has the opposite effect, and people aren’t always transparent. So it just makes me happy to see what you all are building with it. I want to focus on the technical before we come back out to the people part of this. Why’d you choose GPT-2 as a core engine for this? What thinking went into that?

Yeah, so that’s a great question. Just a bit of background on this topic. GPT-2 is one of many models that they’re called large language models, and large language models are a class of models, let’s say, that, one, are really large, they usually have over a billion parameters, and they’re trained on a large training set. So GPT-2 in particular was trained on 8 million web pages, or 40 gigabytes of internet text. And the purpose of these large language models such as GPT-2 is so that once they’re trained on this large training data set, AI practitioners can then apply it to their own use cases. So getting to your question, why did we choose GPT-2, we actually tested several methods and several large language models. We knew beforehand that each had their own strengths and weaknesses, and through that experimentation and iteration we found that GPT-2 was able to both create realistic responses but also realistic responses throughout a conversation, and with sufficiently long responses that mimicked what we see in actual crisis conversations. So to reiterate, the use of GPT-2 came through a series of experimentation and iterations, and with that, with our team, we worked to find what worked best for us, and that’s how we came about with GPT-2.

Yeah, sounds like a solid choice. I know some of the projects, it can be a little hard to understand the quality and what they do, they’re a little black box, so I understand the thinking that went into it. You all must have done your homework to get there.

So it feels like, with the project itself, do you feel like there’s an opportunity to educate developers or even non-developers in the LGBTQ community about positive uses of AI? Because I think we hear a lot of scary things about artificial intelligence in the press and in movies and other things. So do you think there’s an opportunity to showcase the good that’s possible?

Yeah, I love this topic, and short answer, yes, there’s definitely an opportunity to spread the word and educate others about positive use cases of APIs and AIs such as CCS. So more in detail, since we’ve implemented the Crisis Contact Simulator at Trevor, we’ve been super vocal about our work, so we’ve talked to numerous publications, visual media, podcasts, including the one we’re on right now, and we’ve been very explicit in demonstrating how AI can be used in a positive manner. So this has resulted in a huge outreach initiative from us. And again, like I said before, we’re really proud of what we’ve done, and we definitely want to show that to the world and to encourage others to do the same, and to also help them realize that yes, this is possible and this does have actual benefits. And we really want to help others envision how AI can be used in a positive way.

Yeah, it’s good to hear. I spend a lot of time on this subject with APIs, and in particular really trying to showcase positive use cases, so that people don’t focus on the negative that we’re bombarded with in the space. But one of the trends that I’m seeing evolve is around product management and AI projects and API projects being treated as products, not just because we’re selling them, but because there’s a need to focus on the human and ground them more in the human part of doing this, rather than focusing on just AI because it’s cool, or APIs because it’s fun to build an API or whatever. How does this actually solve a human need or a problem? So can you share a little bit about the product management process, how you build this and how you’re going to iterate upon it moving forward?

Yeah, so I want to first start out by saying that our AI work is led by a diverse group of staff, including AI experts and engineers, product managers, software engineers, and so on and so forth. And definitely the product management process is a vital part of that. So, also speaking as an engineer, to me our product management process is really important because it really helps understand and come to a solution for stakeholder needs. So really stakeholders are at the center of our product management process. We really try to develop a close relationship with stakeholders, so this would be different teams throughout the organization, like the training team, a team helping improve the quality of our conversations, so on and so forth. So we develop a close relationship with these teams so that we can really understand, okay, what are the challenges that they’re facing right now, what are they thinking about, what do they really want to solve right now? And our role as an AI team is to really think of solutions which include AI that can perhaps alleviate a lot of these challenges that they communicated to us. And I think, like I said before, this requires a cross-collaborative process. It requires our product managers to have this close relationship, to have a strong communication bridge to understand what’s the situation, what kind of solutions do the stakeholders see as feasible. AI engineers such as myself would perform literature reviews to see, okay, what’s out there, what are the spaces of possible solutions. We have software engineering to help understand the backbone that acts as the foundation for the products that we build. And we also do UX research to really understand more about what people are thinking and to get granular in that thought process. So I think it’s really cool that all these different components are in play and come together in our development process at Trevor.

Yeah, and those characteristics, I would say, is what I see across the healthiest technology projects out there. They bring that human piece to the table and make sure we’re solving a human problem, because otherwise the tech doesn’t matter in any of this. So that feedback loop with all the stakeholders, I’m assuming you’re going to continue using this process and this feedback loop to move forward and iterate on the project. It’s not done, right?

Yeah, for sure, it’s not done. And definitely the feedback process has played a vital part in iterating and improving the Crisis Contact Simulator, which I would say is one of the things on our roadmap. So currently we are actively gathering feedback from the digital counselors who use the Crisis Contact Simulator to practice their skills, and we actually did have an iteration in which we used that feedback to improve the Crisis Contact Simulator. So in the first iteration of the Crisis Contact Simulator, we actually got feedback that the AI chatbot wasn’t answering all that well in very specific cases. So while these cases didn’t represent the majority of cases, the cases that we got feedback on definitely represented room for improvement for us. So the second iteration’s goal was to expand the usage of the Crisis Contact Simulator, but we also took that opportunity to improve upon it based on the feedback that we got. And since implementing those changes, we’ve definitely been seeing a lot less of that type of feedback coming in. And I think that’s one of the ways in which we map out, okay, what are the next steps that we need to do for CCS.

Yeah, so important. So I noticed throughout this, you’re here telling the story of this, transparency seems to play a pretty significant role in how you build technology. Can you talk to me about the role that transparency plays in this?

Yeah, so I agree with you, transparency does play a big role in our projects. So for CCS in particular, I think this manifests in a variety of ways. So for one, we’ve maintained transparency with our stakeholders throughout the entire development process and even post-development. After we’ve deployed CCS, we’ve maintained close contact with the training team, so we’ve been in communication with them about, okay, what are your needs, what does the conversation need to look like, how are we going to get feedback on this, and in addition to that, what are we doing to develop this AI chatbot and what are we doing to evaluate it. So there’s a conscious effort to maintain transparency with our stakeholders. With our stakeholders, I think there’s also transparency with users. So before people engage in the conversation with CCS, we actually let them know that, oh, you’re talking to an AI chatbot, just so that they know that, okay, this is a safe space to practice the skills that I learned, and it’s not an actual LGBTQ youth, nor is it an actual instructor that I’m talking to. So I think transparency has played a role in those two ways for this project.

Yeah, it’s powerful, and I feel it’s a model, not just for other AI projects to follow, but also, for me, APIs, because what you describe, your product management process, the feedback loop, this transparency, I feel like this kind of disclosure is all critical for more human-centered technology. This is how we’re going to make it better serve us and not create dark spots or black boxes or things that make it hard for us to understand what technology does. So this is really important to me, I think, to share with our audience.

So how is CCS funded, how does the project sustain itself?

Yeah, so that’s a great question. So the way it was initially funded was that in 2019, the Trevor Project, our organization, we applied for Google’s AI Impact Challenge, and we won that challenge through our proposal. And through that challenge we won, we were allotted 2.7 million dollars in grants. And like I mentioned before, we actually received full-time support from Google.org fellows who we worked alongside for half a year for this project, and for the two other AI initiatives that I briefly mentioned. And these Google fellows, they ranged from not just software engineers but UX researchers, product specialists, and program managers, and they, in addition to the funds that we received, really helped us build a really strong initial project that we can iterate off of.

Yeah, it’s funding like that, in grants and programs, are important. The money’s important, but that kind of support I think is pretty critical for the success of these types of projects. What does the roadmap look like, what’s next for CCS as far as serving the community?

It’s a really exciting question. So the two things that we’re looking to do is, one, we’re definitely looking to see how we can find more use cases for CCS within the Trevor Project. So like I said before, CCS is part of our training curriculum, so we’re really on the lookout, through conversations with stakeholders, for different ways that we can use CCS to support our counselors. Would it help to use CCS to refresh counselor skills after they’ve been taking conversations for a while? Would it help to reinforce specific skills that we just developed and that we now want them to learn, so on and so forth? So we’ve been talking to our stakeholders, maintaining communication with them, and again, on the lookout for these types of use cases throughout the Trevor Project. And the second thing is that we’re also looking to use CCS to help other organizations outside of the Trevor Project. So we’ve been talking to a number of organizations that may find a benefit and value in using CCS, and seeing what their needs are and if CCS would help them, and if so, what that iteration of CCS would look like. And that’s been really interesting conversations, because that also loops back to the topic about educating others about the positive use cases of API. I think really providing people with this AI chatbot would play a huge part in that. And so definitely I want to say that, if you the audience has resonated with anything that I said or sees any benefit from using CCS, we’re definitely looking for organizations that would want to engage in that conversation.

Yeah, I think that’s where the power is really going to come, and the impact’s going to just exponentially grow. Because for me, like you said, AI is just a tool, APIs are just a tool, and I’m the first to admit APIs aren’t actually the thing that matter, it’s the connections, it’s the relationships, it’s the partnerships and the alliances that it brings. So I think that’s good to hear, that that’s a big part of your roadmap and your whole product management process, your feedback loops, your approach to transparent AI, ethical AI practices. I think that’s a model that will translate well for other organizations that want to get involved. So that’s one way other organizations can get involved, but what are other ways that individuals or enterprise or startups can get involved and help with what y’all are building?

Yeah, so definitely, thank you for that call-out. So individuals can help and get more involved with the Trevor Project in three ways: donate, volunteer, or join my team. So for donating, people’s donations will help us in the multitude of initiatives that the Trevor Project has been and will be continuing. So this includes our advocacy efforts, training more and reaching more counselors and reaching more LGBTQ youth. It would also contribute to educating organizations through our public training that we provide. And our work involves a life-saving mission that really helps LGBTQ youth with the crisis that they’re facing.

So secondly, individuals can volunteer. Volunteers are the backbone of our crisis services, and they are a large part of the team that talks to LGBTQ youth that reach out to us. And volunteers receive extensive training, there’s like a 40-hour virtual training, the practice role plays, and additional support that we provide to volunteers. So if you’re thinking about volunteering, definitely know that you won’t be alone in these conversations. And lastly, we’re growing, we’re hiring a lot. We’re looking for a diverse set of candidates throughout our team, which is a US national team, so people can work remotely. And all these things, donating, volunteering, and joining our team, you can find them on our website, www.thetrevorproject.org.

And we’ll share links in the show notes to help folks get there. But a great cause, very important, but also the technology really leading, leading-edge, progressive approaches to technology. So I think folks can get involved for a variety of reasons and volunteer or help contribute to the work. And I would put out there to all the enterprise organizations and startups, definitely jump in and help sponsor and fund this work, because it’s definitely important.

Well, I have one last question, and if you’ll indulge me, it wasn’t on the list that I gave you, but does Simon, your cat, dictate your internet time at any point during the day? Because I’m a dog person, so my dog Poppy tends to, by the end of the day, she tells me to get off the internet, but some of my coworkers, they’re cat people, and their cat seems to dominate and dictate their internet time. Does Simon do that for you?

Yeah, so I’m nodding because that’s a hard yes. So my computer setup is that my laptop is closed in front of me, and I have it connected to these external monitors, so Simon lies down on my laptop because it’s a heat source. So he’s always kind of monitoring my internet time. And it’s kind of nice, because he lets me know, hey, you’ve been on the internet for a while, let’s hang out. And he’s a very, someone has commented that he’s a very gregarious cat, and I completely agree with that. Like, he loves hanging out with people, so he won’t be shy about giving me a tap on the shoulder. And yeah, that’s something that I really like about him.

It’s good, it shows that he really cares about you and has your best interests in mind. And I think, for me, my dog Poppy, she definitely takes care of me and looks out after me, and I need that, because I’m not always the best at being an advocate for myself. So it’s good to hear that you have Simon doing that for you.

Yeah, well, thank you. I really appreciate you coming here today and sharing your story. This has been very enlightening. I think hearing about the Crisis Contact Simulator has given me some renewed energy when it comes to AI. So I’m impressed with what y’all are building, and I appreciate you coming by and sharing.

Yeah, thank you, and for sure, thank you for this conversation and giving me the opportunity to share our work. Yeah, great. Well, enjoy the rest of your day, and appreciate you being here. Thanks, you too.

Thanks again to Taki for stopping by. For more on Taki, you can find them on LinkedIn, and the Trevor Project at thetrevorproject.org. You can subscribe to the Breaking Changes podcast at postman.com/events/breaking-changes. I’m your host, Kin Lane, and until next time, cheers.