Skip to content

// Industry

Real-time AI for telemedicine

Video visits that connect the first time and stay up, and clinical answers that cite their source — the two places a telehealth product actually fails.

// The problem

Why this is hard

A consult that fails to connect is not a bug report. It is a missed appointment with a clinician's time already spent, and the failures cluster in specific, findable places: a codec profile a desktop browser accepts and an iPhone renders as a black rectangle, audio that degrades past intelligibility long before the video looks bad, a reconnect that never fires.

Alongside it sits the other risk — an answer that sounds right and is not supported by anything. Both problems have the same discipline behind them: measure it end to end, and prove the result instead of assuming it.

  • A permission with a long memory

    A denied camera prompt is remembered per site, so the next request is answered instantly and no dialog opens. The patient waits at a dead preview, and the visit is filed as a no-show.

  • The route left standing

    Where a network blocks UDP, negotiation settles on the relayed TCP path — the last candidate standing. It connects, so the visit counts as a success — while on that leg a lost packet is resent in order, and everything behind it waits.

  • A recording that plays fine

    The recorder is one more subscriber, so a track it never received is absent from a file that still plays. Nothing errors, and the hole turns up months later, when the file is the only copy left.

  • A wait with no clock

    The offer is held open on a promise that only the answer resolves, and nothing in the layer sets a deadline of its own. If the far side never replies, the request waits as long as the process does — and the screen says connecting until someone gives up.

// What matters here

The capabilities that move the needle

// Deep dives

Going deeper for telemedicine

// Proof

Representative outcome

HealthtechClinical RAG

Illustrative — an anonymised, representative engagement; figures are indicative, not a verified client metric.

hallucinated citations in eval

0

on the citation eval set

source-linked answers

100%

every shipped answer

// FAQ

Common questions

Almost always a codec profile or level the iOS decoder will not accept. Safari is strict where Chrome is permissive, so a stream that plays everywhere else renders as a black rectangle on an iPhone with no error anywhere to explain it. It is diagnosable from the negotiated SDP and the bitstream itself, and it is fixable at the encoder rather than in the app.

Answers are composed only from retrieved sources with citations, and an eval suite scores faithfulness on every change — an unsupported answer fails the gate rather than reaching a clinician. Ambiguous cases escalate to a person with the retrieved context attached.

That is a separate engagement from the video itself, and it is the right way to think about it: consent captured and provable, redaction applied on the live pipeline rather than in a batch job afterwards, retention schedules that actually expire, and an immutable log of who accessed what. Storage is the easy part; the obligations are the work.

// Related

Related industries

All industries

// Let's build

Building AI for telemedicine?

Tell us where you are. We reply within a day with a concrete next step.