FaceAge Web Application

Clinical Research Platform | MGB Production: faceage.bwh.harvard.edu

Designed and built a full-stack clinical research web application enabling IRB-approved studies to collect facial photographs from participants and estimate biological age using the FAHR-Face deep learning model. The system is deployed in production at MGB under two IRB protocols (healthy volunteer and MGB patient arms).

Key responsibilities and technical contributions:

• Full-stack development using Python/Flask (Gunicorn, 5 gevent workers), with Jinja2 templating, Bootstrap front-end, and SQLAlchemy/SQLite for authentication data
• Deep learning inference pipeline: deployed FAHR-Face — a 12-layer Vision Transformer (ViT) with a linear regression head on the [CLS] token, operating on 112×112 aligned face crops; weights loaded at runtime from a volume-mounted checkpoint; RetinaFace used for face detection and alignment; per-session median age computed across multiple images and age delta calculated against chronological DOB; results returned to participants asynchronously by email
• Asynchronous task architecture: designed a Redis Queue (RQ) system with a scheduler container polling REDCap every 60 seconds, enqueuing face-age jobs picked up by a dedicated worker container
• Third-party integrations: REDCap API (participant records, questionnaire data, results), Dropbox API (photo storage and file request links), MGB internal SMTP relay (automated transactional emails)
• Authentication and security: implemented full participant auth flow (bcrypt, token-based account activation, password reset, session management, account lockout); added CSRF protection (Flask-WTF), security headers (Flask-Talisman/HSTS), and Redis distributed locking to prevent race conditions on concurrent registrations
• Containerised deployment: orchestrated 5-container Docker Compose stack (nginx reverse proxy with SSL termination, Flask app, RQ worker, RQ scheduler, Redis); cross-platform multi-arch builds (docker buildx) for Apple Silicon → amd64 Linux server
• E-consent and data capture: built a multi-step participant flow covering study-arm selection, electronic consent with PDF generation, baseline and daily questionnaires, and photo upload (file or webcam)

Tech stack: Python · Flask · Gunicorn · SQLAlchemy · Redis · RQ · Docker · nginx · PyTorch · HuggingFace Transformers (ViT) · RetinaFace · REDCap API · Dropbox API · SQLite · Flask-WTF · Flask-Talisman · Flask-Mail

Underlying model: Haugg F, Lee G, He J, Nürnberg L, Bontempi D, Bitterman DS, Catalano P, Prudente V, Glubokov D, Warrington A, et al. Foundation Artificial Intelligence Models for Health Recognition Using Face Photographs (FAHR-Face). arXiv:2506.14909. Under review at Lancet Digital Health. arXiv