Haven.AI is a proactive AI chatbot designed to detect users' specific triggers in advance and provide real-time emotional support when they need it most.
The Problem
01
Addiction Cycle
Most people trying to quit vaping struggle with unexpected cravings triggered by stress, routines, social settings, or boredom.
03
Emotional Impact
Users feel defeated and doomed after giving in to cravings, creating a negative cycle that's difficult to break.
02
Failed Attempts
Traditional quit vaping apps lack personalization and real-time support, leading to relapse when cravings hit.
04
Health Epidemic
Scientists now link vaping to serious health concerns, meanwhile a growing number of adolescents and young adults are stuck in the addiction cycle while mental health support services remain largely unaccessible.
Our Solution
Haven.AI is a proactive AI-powered chatbot with practical tools to help users quit vaping for good.
​
Unlike traditional quit apps that are passive and generic, Haven.AI proactively identifies potential craving moments and intervenes with personalized and therapy-like support before users relapse.
​
Our solution is the only one that caters to both users who want to quit cold-turkey and those who prefer to gradually reduce their vaping habits.
​
AI Hypothesis
If the AI initiates chat with the HAVEN.AI user to suggest personalized and actionable recommendations or distractions to interrupt the craving spiral, with +60% precision, he/she is in the position to resist the craving to vape, which in turn improves their life quality in a perceptible way and increases trust with the AI while also increasing subscriptions, positive reviews and referrals, ensuring our company grows in revenue.
Key Features
Haven.AI MVP
As Product Lead, I managed a cross-functional team of eight - including a UI/UX designer, frontend and backend engineers, and a junior product manager - to design, develop, test, and deploy the Haven.AI MVP within 10 weeks. The beta version has been launched through social media channels and is currently in the hands of early users, allowing us to collect real-world feedback and rapidly iterate on core features. If you’re curious to see how it works, you can download the app and try it out - we'd love your feedback 😀
Registration & Onboarding Quiz within the chat
Home Screen for users who wish to quit cold-turkey (based on onboarding quiz responses)
Home Screen for users who wish to quit vaping gradually (based on onboarding quiz responses)
Craving/Vape Session Logs and Proactive Check-in
Self-Help Toolbox
User Research
As the Product Lead, I created the Voice of Customers interview script with 16 open-ended questions focusing on users' vaping habits and emotions associated with their quitting journey, knowledge, trust and experience with competitor mobile apps and AI chatbots, features they believe would be helpful and willingness to pay. Over 2 weeks, we interviewed 18 people, including direct users and Wellness/HR professionals. Here are the top 3 insights we learned from the research:
1
Anxiety & Stress were the main triggers
Many vapers either started or continue to vape due to anxiety or stress. The habit has now become automatic.
2
Feeling Defeated by the Vape
Users expressed a general feeling of being “defeated by the vape.” Many users have tried many times to quit with no success and are now worried that they will be an addict forever.
3
Users want a Balanced Intervention
The idea of a proactive chatbot is great but too many notifications would be annoying. Users are also interested in support tools and gamified features.
Customer Personas
Based on our research and early interviews, we identified several primary customer segments including Gen Z and millennial vape users (aged 16–35), young professionals trying to quit for mental or physical health reasons, and users with a history of stress-related nicotine dependence (including cigarette use).
We also identified secondary customer segments, such as HR professionals and Wellness Directors at organizations and schools.
​
We chose to focus our MVP on the customer segment of young professionals and university students because they represent a high-need, emotionally driven group with a clear problem to solve: breaking automatic craving cycles and managing emotional triggers like stress, anxiety and social pressure.
​
This audience is often juggling work, deadlines, school pressures, and social expectations - leading to unconscious vaping and frustration with relapse. They need a supportive, real-time solution that fits seamlessly into their day, helps them build emotional regulation skills, and reduces cravings without judgment. By targeting this segment first, we can clearly show how Haven AI’s proactive support and personalized chatbot can build trust, reduce usage, and form healthy new habits - paving the way for broader adoption across other customer segments and B2B partnerships.
Customer Journey
AI Model Selection
We chose GPT-4 Turbo as the LLM for Haven AI's chatbot for the following reasons:
​
-
Faster Response Time – GPT-4 Turbo delivers quicker outputs, which enhances the overall user experience during live interactions.
-
Strong Performance – Its responses were consistently high quality, especially in terms of empathy and relevance, which is crucial for our therapeutic use case.
-
Seamless Integration – GPT-4 Turbo is part of OpenAI's broader developer ecosystem, designed to plug easily into full-stack product workflows. This reduces engineering complexity and enables lean teams to move fast.
During the model selection process, we also tested Meta's LLaMA 3.1. We discovered the following differences, which led us to ultimately choose GPT 4.0:
AI Architecture
This is the core architecture powering Haven AI’s conversational engine.
​
When a user sends a message, Haven retrieves the last 10 message pairs to maintain contextual continuity. These messages are included in the prompt, which is further enriched using semantic retrieval through a Retrieval-Augmented Generation (RAG) pipeline. The RAG system queries a vector database built on cognitive behavioral therapy (CBT) and CounselChat datasets to enhance the relevance and emotional intelligence of the conversation.
​
GPT then uses the constructed prompt to generate a personalized response. The full conversation history is stored in MongoDB, while only the most recent 10 messages are used to inform future prompts to ensure efficiency and maintain coherent dialogue flow.
​
Haven also includes a proactive check-in layer. Using APScheduler, the system automatically triggers the bot to reference each user’s check-in preferences. GPT initiates a check-in by prompting the user to reflect on their progress and log any missed cravings or vape sessions directly within the chat.
​
As the conversation unfolds, a LangChain agent runs in the back-end to detect unreported cravings or usage events based on the user’s natural language. These are automatically captured and stored in the database, enabling GPT to track behavioral patterns and support users with increasingly personalized, data-informed responses without the user needing to leave the chat.
Data Pipeline/Strategy
Step 1 - Data Identification
We sourced two datasets from Hugging Face:
1. CBT (Cognitive Behavioral Therapy) dataset – This provided therapy-style conversations, useful for modeling structured therapeutic dialogue.
2. Counsel Chat dataset – This included real-world counseling Q&A across mental health topics.
Step 2 - Data Processing
Data Cleaning: Removed duplicates, low-quality responses, and irrelevant content.
Standardization: Unified formatting and structure of both datasets for consistency.
Step 3 - AI Model Preprocessing
We combined the two datasets into a unified knowledge base by embedding the preprocessed data using OpenAI’s embedding model to support our Retrieval-Augmented Generation (RAG) setup.
We then indexed the embedded data using Pinecone vector store to enable fast and relevant retrieval during user interaction.
Step 4 - Continuous Feedback
Round 1: Plain GPT-4 Turbo model, no RAG. The responses lacked context and empathy.
Round 2: Introduced RAG with both datasets. Huge improvement in relevance and tone.
Rounds 3-6: Focused on prompt engineering to shape Haven’s voice.
We updated the prompt four times, each version tuning the tone to be more empathetic, supportive, and therapeutic, like a friend-therapist hybrid. We used qualitative feedback and internal test conversations to evaluate the tone, emotional intelligence, and helpfulness of responses. The final prompt was chosen based on the best alignment with Haven's brand voice and user support goals.
User Stories & Acceptance Criteria
The table below includes the prioritized list of user stories for the MVP:
Challenges & Lessons Learned
Making the Chatbot Feel Human
To craft responses that felt genuinely empathetic and emotionally intelligent, rather than robotic or generic, we used a combination of CounselChat + CBT trained dataset to create a RAG for the response to be more efficient, empathetic and humane.
Feature Prioritization
Ambitious feature ideas outpaced our frontend capacity with only two engineers so we prioritised high-value features and sharpened the product focus.
Building Conversational Memory
Maintaining meaningful context in conversations was a challenge so we implemented a lightweight memory system where the last 10 user-chatbot message pairs are continuously fed into the model to personalize responses, and make interactions feel more natural and coherent.
Seamless Logging
Enabling automatic logging of cravings and vaping data during natural conversation without disrupting the user experience. We implemented an agent-tool framework where the model identifies relevant information in real-time chat and triggers backend functions seamlessly within the flow of conversation.
Balancing GPT-4 Power with Performance
Long, detailed prompts improved response quality for addiction topics but slowed response time. Through many iterations and testing rounds, we optimized prompt structure to balance empathy and brevity, learning that effective prompt engineering is not just about what the model understands, but also how fast it can respond in real-time support scenarios.
Product Roadmap
Phase 05
Phase 01
MVP Launch
-
Therapy trained chatbot
-
Seamless logging within chat
-
Manual logging
-
Gamified progress
-
Support tools
Enhanced Personalization
-
Deeper logging analytics
-
Calendar and smartwatch integration
-
Increased GPT memory and premium offering.
Community Engagement
-
Implementing secure, moderated peer-to-peer forums and group support features.
Mental Health Therapy
-
Train LLM for adaptive therapy paths beyond addiction
-
Connect users to human therapists.
Enterprise Licensing
-
B2B dashboard for schools, universities, therapists and HR departments for employee wellness initiatives.
AI
Scalability
Haven AI is built with long-term AI scalability in mind, ensuring our chatbot continues to deliver effective, personalized support as we gather more behavioral and usage data.
We have implemented prompt optimization and real-time context management to adapt to evolving user behavior. This allows Haven AI to maintain relevance and reliability, even as patterns of craving, relapse, or motivation shift over time.
To reduce the risk of hallucinations and minimize AI bias, we fine-tune our RAG layer using diverse mental health and CBT-based datasets, while continuously monitoring outputs for safety and unintended responses. This scalable architecture enables Haven AI to provide emotionally intelligent, context-aware support that grows more accurate and personalized as user needs evolve.