Inference as a Service lets users run inference on remote pre-trained models. This convenience comes at a cost: to obtain a prediction, the user must provide an input in the clear. In sensitive domains (e.g. healthcare), some information cannot be disclosed: regulatory frameworks such as the GDPR make the casual exposure of personal data a liability rather than a convenience. The result is a paradox: the service exists, but it should not be used due to privacy issues. Fully Homomorphic Encryption over the Torus (TFHE) offers a way out: letting the server compute directly on encrypted inputs, so that the user's data, every intermediate value, and the final prediction all remain encrypted end to end.