Aletheia

AI-text detection API. Runs locally or on Modal.

Run

Local (needs a CUDA GPU):

uv sync
uv run app.py

Modal:

modal deploy app.py

API

GET  /health   service status, model list
GET  /models   description and token window per model
POST /detect   { "text": "...", "models": ["fakespot"], "verbose": false }

models is optional; omit to run all. verbose is optional.

Score: P(AI) ∈ [0, 1]; >= 0.5 means AI Generated. Each result includes score, score_std, and windows; set verbose=true to receive per-window details. Text must be 40-5000 whitespace-split words.

Add a detector

1. Subclass BaseDetector in detectors.py.
2. Add a Model(...) row to MODELS in app.py.
3. Append the class to load_trained in app.py.

Layout

app.py          API, engines, Modal services
detectors.py    trained detectors
binoculars.py   Binoculars detector