app/render.yaml

# Render Blueprint — deploys the Gridlock demo as a single Docker web service.
#
# Usage:
#   1. Push this repo to GitHub/GitLab.
#   2. In Render: New + > Blueprint > pick this repo. Render reads this file.
#   3. Deploy. The image builds the frontend and serves it with the API on $PORT.
#
# NOTE ON PLAN: the image bundles PyTorch + a sentence-transformer + three
# gradient-boosting libraries. Inference needs ~1.5–2 GB RAM, so the free 512 MB
# instance will OOM on the first /api/predict. Use the Standard plan (2 GB).
services:
  - type: web
    name: gridlock-demo
    runtime: docker
    plan: standard            # 2 GB RAM; "free"/"starter" (512 MB) will OOM on predict
    dockerfilePath: Dockerfile
    dockerContext: .          # build from repo root so src/ + models/ are in context
    healthCheckPath: /api/health
    autoDeploy: true
    envVars:
      - key: PORT
        value: "8000"
      # Model is baked into the image; never reach out to Hugging Face at runtime.
      - key: HF_HUB_OFFLINE
        value: "1"
      - key: TRANSFORMERS_OFFLINE
        value: "1"
      # Keep tokenizer threads predictable on small instances.
      - key: TOKENIZERS_PARALLELISM
        value: "false"