Discover how I designed a GenAI-powered chatbot that offers therapeutic conversations and personalized poetry to support New Yorkers facing mental health challenges. Built as part of the 2025 GenAI Intensive Capstone, this project explores how AI can provide free, anonymous, and emotionally resonant support—especially for those underserved by traditional care systems. From RAG-based document understanding to real-time weather grounding and emergency lookups, the prototype blends technical innovation with a human-centered mission.

Mental health in New York City is in crisis. As a resident and tech enthusiast, I wanted to explore whether Generative AI could play a role in addressing this urgent issue. That curiosity became the foundation for my GenAI Intensive Capstone Project: The AI Poet Therapist for New Yorkers—a chatbot designed not just to talk, but to understand, uplift, and support through conversation and poetry.

This blog post shares the story behind the idea, the real-world problems it aims to address, the technology that powers it, the challenges I faced building it, and a glimpse into what the future could hold.

🌆 The Real Problem: Mental Health in NYC

New York has always been a city of extremes. Energy. Diversity. Ambition. But behind the bright lights, many New Yorkers are quietly suffering.

Since the COVID-19 pandemic, mental health issues have surged—especially among children, teens, LGBTQ+ communities, and racial minorities. Anxiety, depression, and burnout are common. Many residents lack access to affordable care, and stigma still prevents too many from reaching out for help. Services are often underfunded, overburdened, or simply not designed for the populations who need them most.

As city leaders call for a community-centered, equity-first response, I wondered: Could an AI agent play a supportive role in this ecosystem?

🤖 How GenAI Can Help: AI Companion That Listens

The core idea behind my application is simple but powerful: create an AI chatbot that anyone in New York can talk to—any time, anywhere, in privacy, without judgment and always free.

Using Generative AI, the chatbot would engage users in therapeutic conversations, offering emotional support and coping tools. More than just a chatbot, it would also:

• Listen actively and identify signs of distress

• Suggest grounding exercises or mindfulness techniques

• Connect users with real emergency contacts, when needed

• Generate support materials (sound, image, poem, song), tailored to the person’s emotional state

• Support multiple languages for inclusivity

• Offer mobile access, making it available on-the-go

In essence, I wanted to prototype an AI that feels a little more human. A bit more poetic. And a lot more caring.

💡 The Concept: Therapy Meets Creativity

While the AI agent provides support through dialogue, it also offers a creative twist—at the end of a conversation, it generates a personalized poem based on:

• The user's emotional state

• The current weather in NYC

• A famous singer associated with the city

Why poetry? Because sometimes, when words fail, art speaks. Art do not need capitalism. And also because making a song is too complicated with my python level.

By combining therapeutic conversation with a lyrical, almost artistic gesture, I hoped to offer not just comfort—but also a spark of joy, a moment of beauty, and a deeper connection with the cultural soul of New York.

🛠️ The Tech Stack: Building the Prototype

The system was designed using Gemini-Flash-2 from Google for most of the generation tasks. Here's how it all came together:

🧠 Step 1: Understanding New York’s Mental Health Landscape

I processed a PDF report on mental health in New York using RAG (Retrieval-Augmented Generation). The data was embedded and stored so the chatbot could reference it during conversations—bringing contextual knowledge into every answer.

🌤 Step 2: Connecting to the Real World

To make the experience feel current and relevant, I added Google Search Grounding to fetch:

• The current weather in New York

• A list of well-known singers currently living in the city

This gave the chatbot a richer, real-time context to work with.

🚨 Step 3: Safety First — Emergency Lookup

Rather than relying on the LLM to guess emergency numbers (and risk hallucination), I built a small database lookup function. This guarantees accurate and reliable emergency contacts—critical in high-stakes conversations.

🧩 Step 4: AI Agent Design with LangChain and LangGraph

The system uses LangGraph with three nodes:

• Human – the user interface/input

• Chatbot – the LLM-powered conversational agent

• Tool Node – retrieves emergency contact data

While basic in architecture, this modular design makes the system flexible and ready for future enhancements.

⚠️ Limitations (And Lessons Learned)

As someone without a strong coding background, I faced a few hurdles:

• I couldn't fine-tune a model to behave more like a trained therapist. That would have taken deeper ML knowledge and more compute resources.

• I wanted to build a UI (possibly with Gradio or Streamlit), but I abandoned the attempt midway.

• I also tested text-to-speech for the poem feature, but the free Python packages I tried weren’t up to the emotional nuance I wanted. It ended up sounding robotic, not poetic.

Despite these setbacks, I learned a lot—from designing LangGraph workflows to embedding documents and grounding outputs with search. More importantly, I proved that a meaningful GenAI prototype doesn’t require perfect code—just a clear purpose and thoughtful execution.

🔮 What’s Next: Dreaming Bigger

This is just the beginning.

If the technology evolves (and I get better at building things), here’s what I’d love to see in future versions:

• Voice chat support, so users can speak instead of type

• Real-time video analysis, where the AI can interpret facial expressions and emotion

• An AI avatar, for a more human, expressive interaction

• Fine-tuned models, trained on therapist-like dialogues with cultural sensitivity for New Yorkers

• Offline functionality, so access isn’t dependent on stable internet

And maybe—just maybe—it becomes more than a prototype. But always free for sure.

🗽 Final Words

I didn’t build a perfect app. I didn’t even build a full one. But I built something that matters.

The AI Poet Therapist is a small gesture toward a big challenge—mental health support that’s inclusive, accessible, and a little more human. For all the noise in New York, sometimes what someone really needs is someone—or something—that listens.

Thanks to the GenAI Intensive Course for the inspiration, the guidance, and the opportunity to turn this idea into a tangible, working system. Even in its early form, it showed me what’s possible when technology meets empathy.

Want to know more? Watch the video with a demo of the chatbot: