Autism Insight
A full-stack autism-screening prototype — a Next.js/Firebase/MongoDB app for questionnaire-based screening and admin review, paired with a FastAPI + PyTorch service exploring image- and video-based classification. Built as an end-to-end product exercise, not a validated diagnostic tool.
Overview
A full-stack autism-screening prototype — a Next.js/Firebase/MongoDB app for questionnaire-based screening and admin review, paired with a FastAPI + PyTorch service exploring image- and video-based classification. Built as an end-to-end product exercise, not a validated diagnostic tool.
Tech Stack
Challenges & Learnings
This project explores what a screening tool looks like as a full product, not a bare model script: user accounts, screening history, and an admin review layer sit on top of a FastAPI prediction backend. The frontend (Next.js, Firebase Auth, MongoDB) protects routes with cookie-based role checks and gives admins filterable, printable analytics. The backend exposes three prediction routes. The questionnaire route (AQ-10-style yes/no items) is currently a rule-based sum-and-threshold classifier rather than a trained model. The image and video routes run a small custom-trained CNN over uploaded face images; for video, it samples frames evenly and takes a majority vote across per-frame predictions, with no temporal modeling. Real, visible tradeoffs: uploaded files are written to and deleted from disk synchronously per request rather than streamed, and CORS is wide open, both consistent with a prototype deployment rather than a hardened service. Worth stating plainly: this is a full-stack engineering exercise around a screening workflow, not a clinically validated or diagnostic tool — face- and video-based autism prediction in particular has no established clinical basis, and the project should be read as a technical case study in building the product around that idea, not a claim about real-world accuracy.