Inspiration

SerenAI was born out of a deep desire to make mental health support more human, more accessible, and more insightful—without losing the soul in the science.

I noticed that most chatbot solutions in mental health simply give canned advice or mimic empathy. But what if an AI assistant could go beyond surface-level support? What if it could understand not just the words, but the emotions, the trends, and even the urgency—and present that insight to both the user and the therapist?

I personally witnessed the silence many people feel—too scared to open up, too confused to explain what they’re going through. I wanted SerenAI to be a safe space to start that conversation, and a bridge to the therapist on the other side who wants to help.

What it does

SerenAI is a soothing mental health chatbot that lets users talk about their feelings in a calm, welcoming interface—and does a lot more under the hood:

-🧠 Emotion detection after every message using nomic-embed-text + an emotion classification model.

-🚨 Crisis risk scoring to determine if a message needs urgent attention (NORMAL / HIGH RISK / CRISIS).

-🪄 Real-time recommendations (books, movies, songs) tailored to the user’s emotions—complete with icons and categories!

-🧸 A calming bear companion animation and beautiful chat interface.

-📊 On the therapist side, a dashboard shows all users, their latest messages, emotional tags, and a live trend graph showing risk scores and emotional progression.

  • 🚩 If a CRISIS message is detected, the therapist is immediately alerted with a red card and call-to-action prompt.

How we built it

Frontend:

I built the frontend using React + Vite with calming, elegant UI inspired by mental health apps. We used:

-Lottie animations for mascots (like the bear!) -Framer Motion and CSS transitions for soothing effects -Chart.js for emotional trend graphs -Auth0 for authentication between User and Therapist views

Backend:

We used FastAPI and integrated:

-Gemini API (Google) for: -Emotion-related recommendations -Clustering emotional phrases -Risk scoring -Ollama (nomic-embed-text) for phrase embeddings -Custom emotion classifier using bhadresh-savani/distilbert-base-uncased-emotion -CSV logs per session with name, email, message, emotions, and risk score -Google crawler to enrich recommendation visuals (books, movies, music)

Architecture:

We kept the backend modular:

  • "/chat" handles conversation

  • "/predict" for emotion detection

  • "/crisis" for scoring

  • "/recommendations" for content

  • "/logs" for therapist access

Challenges we ran into

  • Emotion classification isn't always perfect—so we trained a fallback model and carefully tuned the prompts to be safer.

  • Frontend performance—with live recommendations and graph rendering, we had to optimize layout and rerenders.

  • Data formatting—making sure we stored emotions in a way that the chart could display them across multiple sessions was tricky.

  • Crisis detection—we had to be very sensitive with thresholds and responses so the system doesn’t either over- or under-alert therapists.

Accomplishments that we're proud of

  • We built a functioning full-stack application in under 48 hours with real AI features, not just mockups.

  • The therapist dashboard provides real emotional insight in a beautiful UI.

  • Crisis alerting works and is logged for traceability.

  • Our chatbot feels comforting, not robotic.

  • We pulled off agent-like behavior with Gemini's reasoning capabilities, summarizing emotional clusters and guiding therapists.

What we learned

  • How to use Gemini prompts for psychological reasoning and emotional clustering.

  • How to balance UX design with emotional sensitivity—users dealing with mental health issues need clarity, calm, and comfort.

  • Building agentic AI systems isn’t just about tech—it’s about giving meaning and insight to data.

  • We saw the power of multi-modal AI systems, where text, emotion, reasoning, and visualization come together to help real people.

What's next for SerenAI

  • Live crisis alerts on the therapist dashboard with a call button and real-time notifications.

  • Voice conversations so users can talk to SerenAI like a friend—and receive responses in audio using speech APIs.

  • Agentic explanations for therapists: Why is this emotion graph trending upward? SerenAI will soon explain the emotional clusters, changes, and risk progression with Gemini-powered summaries.

  • User history timeline, so therapists can see past sessions and mood swings over time.

  • Integration with licensed psychologists for hybrid human + AI intervention.

Built With

Share this project:

Updates