-
-
example 1
-
eval 1
-
example 2
-
eval 2
-
example 3
-
eval 3
Inspiration
Mental health is one of the biggest global challenges today. Millions of people avoid therapy due to stigma, high costs, or lack of professionals. We were inspired to design an AI system that can act as a safe, accessible, and empathetic ally. Instead of a static chatbot, we envisioned a dynamic companion that adapts its therapeutic style to what the user really needs in the moment.
What it does
EmpathAI is an AI-powered psychological chatbot that dynamically switches between five therapeutic approaches:
- Cognitive Behavioral Therapy (CBT)
- Person-Centered Therapy
- Psychodynamic Therapy
- Solution-Focused Brief Therapy (SFBT)
- Dialectical Behavior Therapy (DBT)
It uses an LLM pipeline where one model classifies intent, another generates a therapy-style response, and a third independently evaluates it for empathy, safety, and alignment.
How we built it
Designed a multi-LLM architecture with three components:
- Intent Classifier → Detects therapeutic need from user input.
- Main Chatbot → Generates the response in the chosen style.
- Evaluator → Scores empathy, listening, guidance, and ensures (100\%) ethical compliance.
- Intent Classifier → Detects therapeutic need from user input.
Crafted style-specific prompts for each therapy type.
Built a dataset of 250 diverse queries covering real mental health scenarios.
Experimented with different model setups (LLaMA–Gemma combinations) to find the most accurate and balanced configuration.
Challenges we ran into
- Handling ambiguous queries that could fit more than one therapy style.
- Preventing evaluation bias when the same model acted as both classifier and evaluator.
- Working with limited, manually created datasets instead of real-world user logs.
- Ensuring absolute safety: no harmful, judgmental, or unsafe advice.
Accomplishments that we're proud of
- Built a working adaptive chatbot that achieved:
- 95% intent classification accuracy
- Empathy score: 4.2/5
- Active listening: 3.9/5
- Style adherence: 4.9/5
- 100% safety compliance
- 95% intent classification accuracy
- Developed one of the first multi-LLM evaluation frameworks for mental health.
- Showed that AI can move beyond scripted answers to provide context-sensitive, person-centered support.
What we learned
- Mental health language is highly subjective: sometimes multiple styles are equally valid.
- LLM-as-Judge works well, but evaluators must be independent to avoid inflated scores.
- Empathy and tone matter just as much as accuracy—AI must be designed for compassion and responsibility.
- Building safe AI for health requires balancing innovation, ethics, and human-centered design.
What's next for EmpathAI – Dynamic Chatbot for Mental Health Support
- Add multi-turn memory so the chatbot can remember past conversations.
- Use Retrieval-Augmented Generation (RAG) to integrate real mental health resources.
- Incorporate real user feedback to fine-tune therapeutic alignment.
- Partner with mental health professionals for clinical testing.
- Deploy as a mobile/web app to make psychological support accessible to anyone, anywhere.
Log in or sign up for Devpost to join the conversation.