Aven is a next-generation exposure therapy tool designed to help individuals with social anxiety practice difficult conversations in a safe, deeply realistic, and scientifically grounded environment.
By combining graduated exposure therapy simulation with a fine-tuned RoBERTa cognitive distortion classifier and vocal biomarker analysis, Aven offers a private, clinically rigorous space to overcome social fear.
Practice 25 highly distinct, real-world scenarios across 6 core domains: Authority, Strangers, Group Dynamics, Assertiveness, Workplace, and Intimacy.
- Escalating Difficulty: Characters scale from Level 1 (Warm & Cooperative) to Level 5 (Hostile, Interrupting, and Dismissive).
- Persistent Memory: Characters remember your past interactions and organically bring up your previous failures or successes to maintain realism.
- Human-like Latency: The AI simulates human cognitive load, responding instantly when friendly (L1) but pausing deliberately to "judge" you before delivering hostile replies (L5).
Aven's core research contribution is a custom-trained RoBERTa model that operates asynchronously during your session.
- Analyzes 15 distinct cognitive distortions (e.g., Catastrophizing, Mind Reading, All-or-Nothing).
- Features a dual-head architecture to provide multi-label classification and severity scoring (1–5) simultaneously for every message.
Aven listens to how you speak and observes your non-verbal cues in real time.
- Avoidance Detection: Flags hedging, deflection, topic changes, and message abandonment mid-session.
- Vocal Baseline Analysis: In Voice Mode, Aven measures your pitch elevation, jitter, filler word rate, and speech rate against your personal baseline with continuous mic hardware deconfliction.
- Facial Emotion & Tension Tracking: Uses on-device MediaPipe Face Landmarker to extract 52 facial blendshapes, computing a composite Facial Tension Index and real-time emotion breakdowns (Happiness, Sadness, Anger, Fear, Surprise, Disgust).
- Procedural 3D Visualizer: Driven by WebGL, React Three Fiber, and custom shaders with multi-light rigs and audio-reactive particle rings.
- Dynamic State System: Visualizes AI states (Idle, Listening, Speaking, Thinking) with state-aware glowing colors and physics distortion.
- Interactive Voice Greetings: Landing page orb features conversational intent handling and automatic audio greetings.
At the end of a session, Aven generates a beautiful, actionable CBT report:
- Color-coded annotated transcripts highlighting distortions and assertive moments.
- Specific, CBT-grounded reframes using your exact words.
- Assertiveness scoring (1–10) with breakdowns on directness and confidence markers.
- Clinician Portal Sync & Fear Hierarchy Tracking: Live synchronization of session analytics and automatic completion tracking across user fear hierarchies.
Aven prioritizes user safety and seamless streaming performance above all else.
- SUDS Tracking: Mid-session Subjective Units of Distress Scale (SUDS) checks. If SUDS > 75, the app automatically triggers a pause protocol.
- Crisis NLP Detector: A separate, async NLP pipeline constantly monitors for crisis language, surfacing immediate helpline resources and safely terminating the session.
- Integrated Breathing Tools & Ambient Soundscapes: Procedural Web Audio API engine providing brown noise soundscapes and physiological sigh breathing guides.
- Resilient SSE Streaming Engine & Roleplay Guardrails: Stream buffering in React for fragmented JSON packets, combined with real-time
<think>tag sanitization and zero-breakout prompt constraints to enforce 1–3 sentence spoken human realism.
| Feature | Generic AI Chatbots (ChatGPT/Claude) | Traditional VR Exposure Therapy | Aven |
|---|---|---|---|
| Realism | Often break character, use overly helpful "therapy speak", or apologize unprompted. | Scripts are pre-recorded and rigid. You cannot have an organic conversation. | Strictly enforced human realism. Characters never break role, use therapy language, or act like an AI. |
| Analysis Depth | Provide generic, high-level summaries after you ask for them. | Require a human therapist to review recordings later. | Real-time, multi-label distortion detection, facial emotion tracking, and exact-quote reframing. |
| Progression | No structured fear hierarchy or graduated difficulty. | Difficult to tune precisely to the patient's exact breaking point. | 5 precise difficulty levels per scenario, from cooperative to deliberately hostile and gaslighting. |
| Safety Guardrails | Lack physiological monitoring; will continue roleplay even if you are spiraling. | Expensive biometric hardware required. | Built-in SUDS tracking, vocal & facial biomarker stress detection, and crisis termination protocols. |
| Modality | Primarily text-based; voice mode is not optimized for therapeutic latency. | Visual-heavy, but conversation branching is limited. | Voice + Text with <1.2s round-trip latency and mid-session modality switch detection. |
- Frontend: React + Vite, TailwindCSS, Framer Motion, Three.js / React Three Fiber, Web Audio API
- Backend: FastAPI, Python, SQLite (local fallback) / PostgreSQL, SQLAlchemy, Pydantic v2
- Machine Learning & Signal Processing:
- RoBERTa (Transformer): Fine-tuned for multi-label cognitive distortion classification.
- MediaPipe Face Landmarker (CNNs): On-device WebAssembly BlazeFace & Mesh model for extracting 52 facial blendshapes (Facial Tension Index & 6 emotion metrics).
- WebRTC VAD (Gaussian Mixture Models): Voice Activity Detection for speech rate estimation.
- Parselmouth/Praat (DSP): Acoustic signal processing for extracting physiological voice tremors (F0 Pitch, Jitter, Shimmer).
- Generative LLMs & SSE Sanitizer: Dynamic CBT roleplay generation with stream chunk buffering and reasoning tag stripping.
Open a terminal, navigate to the backend folder, and start the Python server:
cd backend
pip install -r requirements.txt
uvicorn main:app --reloadThe backend will run on http://localhost:8080.
Open a second terminal in the root project folder (Aven), install dependencies, and start the development server:
npm install
npm run devThe frontend will run on http://localhost:5173 (or similar).
Disclaimer: Aven is a training tool and CBT simulator. It is not a replacement for professional psychiatric care or human therapists. For the therapist collaboration portal, please ensure your provider is registered.
If you are deploying this project or pushing it to a public GitHub repository, make sure NOT to commit your .env files. The included .gitignore file is configured to exclude .env files and __pycache__ to prevent accidental leakage of your Groq or OpenAI API keys. A template .env.example has been provided for reference.
Copyright © 2026 Aven Social Anxiety CBT Simulator. All rights reserved.
This project is licensed under the MIT License - see the LICENSE file for details.