ED Sentiment Warning API

This project is a mini-app built on Modal that uses Anthropic’s Claude to detect and flag potentially harmful eating-disorder–related captions (e.g. TikTok, Reels, or other social posts).
It builds upon my more extensive college research on semi supervised transfer learning for eating disorder sentiment analysis: Transfer Learning for Eating Disorder Sentiment Analysis The endpoint returns short, supportive warnings similar to the overlays you might see on TikTok, designed to protect viewers and encourage recovery-minded messaging.

Features

• Claude-powered classification
  Uses Anthropic’s Claude 3 Haiku to classify captions as harmful, neutral, or supportive.

• Short supportive warnings
  Returns a very short viewer-facing message (≤15 words), like TikTok’s “This content may promote eating disorders.”

• Autoscaling on Modal
  Deployed as a Modal web function with automatic scaling. Bursty traffic (hundreds of requests) spins up new containers, then scales back down.

• Swagger API docs
  Comes with an auto-generated /docs page for quick testing in the browser.

Tech Stack

• Modal — serverless infra for Python, autoscaling functions
• Anthropic Claude 3 Haiku — LLM inference
• FastAPI — powering the endpoint & docs

Usage

1. Deploy the app

modal deploy app.py

This creates a web endpoint at a URL like:

https://christineastoria--ed-sentiment-warnings-predict-and-recommend.modal.run/

2. Send a request with curl

curl -X POST "https://christineastoria--ed-sentiment-warnings-predict-and-recommend.modal.run/"   -H "Content-Type: application/json"   -d '{"text": "How to not feel like a whale today"}'

3. Example response

{
  "label": "harmful",
  "rationale": "The caption expresses negative body image.",
  "recommendation_for_viewer": "This content may promote eating disorders"
}

4. Interactive docs

Visit:

https://christineastoria--ed-sentiment-warnings-predict-and-recommend.modal.run/docs

to test via Swagger UI.

Load Testing and Autoscaling

To simulate burst traffic, you can send multiple requests:

export URL="https://christineastoria--ed-sentiment-warnings-predict-and-recommend.modal.run/"
export BODY='{"text":"I feel like a whale today"}'

seq 1 200 | xargs -I{} -P 20 curl -s -X POST "$URL"   -H "Content-Type: application/json"   -d "$BODY" >/dev/null

Watch Modal’s dashboard to see containers spin up and scale back down automatically.

Project Structure

functions/ed_sentiment_analysis/app.py     # Modal app definition + Claude integration
README.me  # Project overview and usage

Future Work

• Frontend demo page (instead of Swagger JSON)
• More nuanced labels (e.g. triggering, educational, recovery-focused)
• Support for batch analysis of captions
• Add observability (logging, metrics dashboards)

Why This Project

This project demonstrates how AI infra (Modal) and safe-by-design LLMs (Claude) can be combined to solve a real, sensitive problem: harmful eating disorder content online.
Instead of a toy demo, it focuses on responsible AI use, short warnings, and developer experience.