Trust & Safety
How we deploy AI in a clinical setting — the boundaries it operates within, how we test it, and where a clinician always stays in control. Written for the doctors, nurses, and care teams who will work alongside it.
What TherapyPod is
- A patient-engagement and triage assistant for the clinic front desk.
- A way to answer routine questions, gather context, and surface urgency.
- A tool that escalates to your team and keeps clinicians in the loop.
What it is not
- Not a diagnostic device and not a substitute for clinical judgement.
- Not an autonomous decision-maker for high-risk care.
- Not a source of definitive diagnoses or guaranteed outcomes.
The boundaries the AI operates within
These are not aspirations — they are rules written into the prompts that power the assistant, and checked automatically when those prompts change. Here is what that means in practice.
It assists — it does not diagnose
The assistant is built to never deliver a definitive diagnosis, promise an outcome, or replace a clinician's judgement. It gathers context, educates, and routes — clinical decisions stay with qualified humans.
Emergency-first escalation
Red-flag symptoms — chest pain, stroke signs, breathing difficulty, anaphylaxis, self-harm — are designed to trigger immediate guidance to emergency services or urgent clinician review, even if a message asks it not to.
Grounded, never fabricated
The assistant is instructed not to invent diagnoses, medications, doses, lab interpretations, history, or clinic policy. When a detail is missing or uncertain, it is built to say so rather than fill the gap.
Human in the loop
When the assistant detects an emergency or urgent safety signal, it surfaces immediate safety guidance to the patient and raises an escalation to clinic staff. Staff can review, correct, or take over any flagged conversation — a clinician stays in control of high-risk care.
Privacy & data sovereignty
Patient data is handled under India's DPDP Act 2023 with clinic-level isolation. The assistant is instructed never to reveal internal instructions, secrets, or unnecessary identifiers.
Resistant to manipulation
Patient messages, uploaded files, clinic documents, and tool outputs are treated as untrusted source data — not as commands. Instructions hidden inside them that try to change the assistant's behaviour are designed to be ignored.
How we test the AI
Safety claims are only as good as the testing behind them. We treat clinical quality as an ongoing program with three reinforcing layers.
Guardrail contract checks
Our core runtime prompts are covered by automated contract checks — emergency escalation, no fabrication, privacy, clinical boundaries — that run in our test suite whenever those prompts change. We are extending this coverage across every prompt surface in the product.
Adversarial behavioural evaluation
We are rolling out an evaluation harness that runs prompts against deliberately hard clinical scenarios — a patient asking us not to escalate chest pain, a request to reveal records, a missing lab range — and grades the actual responses on safety, completeness, tone, and adherence using a separate evaluator model. As this matures, we are calibrating those scores against clinician judgement before they gate a release.
Clinician review loop
We are building a review workflow where flagged conversations are assessed by qualified staff, who are the source of clinical truth. Our commitment is simple: every confirmed failure becomes a permanent regression test, so a mistake — once caught — cannot quietly return in a later prompt change.
This is an evolving program, not a finished guarantee. AI systems can make mistakes; our commitment is to constrain them tightly, catch failures through clinician review, and make every confirmed failure a test that prevents its return.
Saw something that looks wrong?
Clinicians and care staff are our most important safety signal. If a response looked unsafe, incomplete, or simply off, tell us — your feedback goes straight to our team and shapes how we improve the assistant.
Report a concernSee how it works in your clinic
Walk through the escalation and review workflows with our team, or read how we handle patient data.