How we help a non-profit project yana
support > 900 people who experience discrimination
Contact us ->
yana is the world's first AI chatbot that supports people who experience discrimination.
Main results
Since the project launch, more than 900 users talked with yana about such a sensitive topic
More than 200 people shared their story with yana and got emotional support and validation
90% of people who saw our intro message sent more than 1 question to yana
Contact us ->
How it started
We started working together in 2022. Back then yana chatbot was just an idea and we were starting completely from scratch. ParsLabs collaborated with a team of conversation designers on this project, our role was technical implementation - training the models, prompt engineering and giving feedback on conversation design.
The project went through two phases: we released an initial MVP in Rasa. Then we made a website and widget update and released a chatbot in VoiceFlow with a new widget. In this case study we will walk you through the process.
Mission
In Germany, there is a lack of help on the topic of discrimination. There are some agencies that offer help but most people don't know how to reach them. Moreover, not everyone feels comfortable discussing their experiences with real humans. There is a lack of centralised information on what you can actually do to deal with your experience, where to get emotional support, how to get active, how to take legal actions, where to find a community of like-minded people. And most importantly, is what happened to you okay (spoiler alert - it's not) and are you alone in this?
This is how the idea of yana got born by Said Haider. Together with Lautmaler agency we helped make this a reality. ParsLabs provided support at different stages of the project, doing AI training, helping with feedback on conversation design, and taking care of the UI.
Meet yana
yana is a German speaking AI powered chatbot that can
support people who experience discrimination.
Vision
Ease barriers
Make it easier for people to get emotional support and get educated
Educate
Mix of buttons to lead the conversation and offer context & LLMs to let people speak
Provide help
Offer help on legal, mental health, practical, psychological, just venting
Features
Smart and empathetic: yana can respond to users in a personalised, empathetic way using data, proofread by our domain experts and psychotherapist.
Knowledgeable: yana is trained on a knowledge base to respond to user questions using LLM grounded on data.
AI-powered & button-based: People can interact with yana in a free text format or using buttons.
Analytics dashboard: ParsLabs built a custom analytics dashboard connected to VoiceFlow to monitor important chatbot metrics.
Custom chatbot widget: ParsLabs built branded chatbot widget to engage website visitors better.
Featured scenarios
Sharing links
Sharing resources as a carousel
Responding with empathy
Evolution of yana
yana undergone two major updates
First MVP was developed in Rasa within 2 months. We wanted to quickly have something working and worked in parallel with conversation design team from Lautmaler. While they were designing we were developing.
This way we quickly realised that Rasa wasn't the right tool for our needs anymore. As our colleagues from conversation designer team have been doing research talking to users they realised that the flows will be more complex than we initially thought and the content we plan to share with our users will be quite extensive.
It was hard to maintain our project in Rasa. Models became too heavy and slow to run, it wasn't the right tool for our conversation design. We needed another tool that could support that amount of data and handle long complex flows. We decided to migrate the project to VoiceFlow.
VoiceFlow was a great choice for this project because it was quite versatile. It combined NLU, buttons, multimedia, LLMs, knowledge base and other features important for this project. It also was great for supporting very long complex flows with multiple turns and conditions.
We combined VoiceFlow features with custom code we wrote in Python and hosted on our secure servers. For instance, we wrote our own code for knowledge base extraction and connected it to VoiceFlow via an API, see the screenshot below.
RAG-pipeline developed using VoiceFlow and self-hosted API
UI is an important part of any chatbot development project.
Before
open-source widget
connected to Rasa
After
custom widget from ParsLabs
connected to VoiceFlow
One of the limitations we encountered while using VoiceFlow is that their standard chatbot widget is quite limited in terms of customisation options. You can't upload your own font, you can change the way input field works, how buttons are displayed, which interactive elements are used.
Our client needed yana to match their brand. That's why yana chose to use a custom chatbot widget developed by ParsLabs, which we connected to VoiceFlow and integrated with their website.
Setting up analytics dashboard
Every chatbot development project has different success metrics.
Advanced analytics dashboard developed by ParsLabs
We needed to measure success and track conversation logs and the analytics dashboard that VoiceFlow offered out of the box wasn't extensive enough. In VoiceFlow analytics you can only see total number of interactions, users and sessions.
For yana, we needed to see what percentage of people who visit the website talk to the chatbot. If the conversion rate was low, if people were visiting the website but not clicking the chatbot icon, it was a sign for us that something needed to change in the way we positioned the widget on the page or it's design.
We also needed to know how many people who see our intro message continue the conversation with yana. If this metric is low it meant that something was off with our intro message, either it was too long, too confusing or just not interesting enough for the user to continue the conversation.
Another metric we were interested in how long are conversations with yana on average on average and how many people start & finish our LLM-powered emotional support flow. While without collecting direct feedback from users it's hard to know if the conversation with the AI Assistant was helpful, this metric allows us to approximately measure success of the conversation. The assumption here is that if users decide to have a long conversation with yana there must be something of value for them.
To track all those metrics ParsLabs created an advanced analytics dashboard for this project, connected with VoiceFlow.
Results
900+
people had a conversation with yana on such a sensitive topic as discrimination
200+
people shared their story with yana and got emotional support and validation
90%
of people who saw the intro message sent more than 1 question to yana
We collaborate together with Lautmaler to continuously update and improve yana, making it more empathetic and knowledgable.
Testimonial
Said Haider,
CEO of yana
Do you find this case interesting? Let's build something extraordinary together. Tell us a bit about your idea and let's discuss how we can help you with our services.