This is a submission for the Gemma 4 Challenge: Build with Gemma 4

What I Built

Mental health crises don't wait for business hours. When someone is experiencing a panic attack at 2 AM, a generic chatbot isn't enough. They need immediate, empathetic, and highly intelligent intervention.

SafeMind AI is a comprehensive emotional sanctuary that bridges the gap between private journaling and active therapy. It acts not just as a conversational chatbot, but as an autonomous therapeutic agent. It proactively identifies emotional distress, maintains persistent long-term semantic memory, and dynamically routes users to standardized clinical tools like Cognitive Behavioral Therapy (CBT) records, PHQ-9 assessments, and grounding exercises.

Key Engineering Features:

• Real-Time Architecture: A hardware-accelerated frontend that talks to a Python/Flask backend, featuring instant DOM updates for mood tracking without page reloads.
• Emergency Kill-Switch: Engineered using JavaScript AbortControllers to instantly sever the LLM network request if a user panics and wants to cancel a thought mid-generation.
• Voice Accessibility: Native Web Speech APIs for real-time dictation, allowing users to review their spoken text before sending it to the model.

Demo

Try the live application here: SafeMind AI Live App
(Note: It may take 30-50 seconds for the Render server to spin up on the first click!)

Here is a complete video walkthrough of SafeMind AI in action:

Code

You can view the full source code and system architecture here:
[https://github.com/shahulhameed-csecore/SafeMind-AI]

How I Used Gemma 4

For this project, I specifically chose the Gemma 4 26B MoE (Mixture-of-Experts) model via API.

While the smaller E2B/E4B models are incredible for edge deployment, SafeMind AI required the deep emotional reasoning capabilities of a larger parameter model to accurately handle crisis intervention and clinical tool routing. However, a dense 31B model could introduce latency, which ruins the conversational therapy experience.

The 26B MoE architecture offered the perfect equilibrium:

1. Agentic Routing: I utilized Gemma 4's advanced instruction-following capabilities to act as a therapeutic agent. If a user expresses overwhelming panic, Gemma 4 autonomously embeds a tool-call in its response, prompting the UI to launch a cinematic "Release (Burn) Exercise" or CBT record directly in the chat.
2. Persistent RAG Memory: To simulate genuine human empathy, SafeMind uses Gemma 4 in tandem with a Pinecone vector database. Past chat logs are vectorized so Gemma 4 can recall if a user was stressed about an exam last week and follow up on it today.
3. High-Speed Empathy: By utilizing the MoE architecture combined with asynchronous Server-Sent Events (SSE), tokens stream instantly to the UI, providing lightning-fast Time-to-First-Token (TTFT) for a seamless voice-therapy experience.