Platform

An engagement layer, not an unconstrained chatbot.

The controls are not bolted on after generation. Redaction, retrieval, scoping and validation are stages in the path a question takes, and a response that fails any of them does not reach the patient.

The path a question takes

Six stages, in this order.

  1. Intake

    The question is redacted before anything else happens

    Inbound text is scanned for personal identifiers and redacted at the edge of the system. The platform is built to avoid holding personal health information at all, so the redaction runs before retrieval, before generation, and before anything is written down.

  2. Retrieval

    Evidence is retrieved before an answer exists

    The question is matched against 28,000+ retrievable passages using vector and keyword search together. Retrieval happens first: the model is answering from a shortlist of real sources rather than producing text and looking for support afterwards.

  3. Scoping

    Each tenant sees its own corpus

    A deployment can be scoped to a defined portfolio, a therapeutic area, or its own private content. Scope is enforced during retrieval, so out-of-scope material cannot reach an answer even if it exists in the shared corpus.

  4. Generation

    Answers are assembled with citations attached

    Answers are written to cite, inline, the passage behind each claim, with the credibility grade of that passage available alongside it.

  5. Validation

    Every generated answer is checked against the guardrails

    As each section completes, a post-generation validator checks it for directive language, diagnosis and assessment, unattributed superlatives and personal health information, and for citations. A section without citations is removed, as is one the validator could not check, and every result is recorded in a guardrail audit log.

  6. Deferral

    What the system should not answer, it defers

    Questions outside the boundary get a deferral rather than an approximate answer, and can be captured in a human-response queue for follow-up where a deployment enables it. Potential adverse-event mentions are flagged on a separate path into pharmacovigilance review.

Built for regulated enterprise deployment

What a deployment gives your teams.

Source-grounded responses

Answers are tied to retrievable evidence rather than unsupported free-form generation.

Adverse-event capture

Potential AE mentions are flagged and routed into pharmacovigilance review workflows.

Deferral, not approximation

Questions it cannot answer are deferred, not approximated — and, where enabled, captured as redacted topics in a human-response queue for source-grounded follow-up.

Training-gap detection

No-result and low-rated questions reveal where content or retrieval needs improvement.

Conversation analytics

Understand what patients ask, where they struggle, and which support topics recur.

Multi-tenant isolation

Separate environments, API keys, webhooks, and feature flags per deployment.

Auditability

Operational logs support review of system activity, usage, and deployment behaviour.

Modular deployment

Ship as a patient experience, a branded support layer, a portal capability, or an integrated service.

Questions we get asked

Straight answers.

What happens between a question and an answer?
The question is redacted for personal identifiers, matched against the corpus within the deployment's scope, answered from the retrieved passages with citations to its sources, and checked against the guardrails, with the result recorded.
Can a deployment change the safety guardrails?
No. A deployment sets its own scope, keys, origins and prompt guidance, but the safety guardrails are set at the platform level and no deployment's configuration can switch them off.

See the pipeline run on your own questions.

Bring the questions your patient-services team actually receives, and watch where the system answers, cites, and declines.

Request a demo Try it live