ABACare Vision

Turn session video into evidence a clinician can check.

After a session, ABACare Vision looks through a consented recording for the moments tied to your client's targets. It proposes each one with a short evidence clip. A clinician accepts, adjusts or rejects it.

The clinician reviews each candidate beside its clip.

Every candidate carries its recorded definition. This is the product, on a fictional clinic with no real video.

Five steps from recording to candidate.

  1. Check the footage

    A quality scan flags dark, covered, or frozen stretches so they are not analyzed.

  2. Scan short windows

    A fast model pass looks at short windows for moments that match the client's active targets and the BCBA's definitions.

  3. Look closer

    A second pass takes a closer look at each proposed moment and keeps what is visible apart from the label it suggests.

  4. Check the rules

    Plain code confirms the label exists, the target belongs to this client, and the times fit the recording. Anything else is dropped.

  5. Cut the evidence

    Each surviving candidate gets a timestamp, a short evidence clip, a confidence, and its provenance: recording hash, model ids, prompt version, and definition version.

The clinician decides. The model only proposes.

  • See the definition behind the suggestion. Inspect the recorded detector definition and evaluation context beside the clip. Review at 1x, 2x, or 4x; a newer detector never replaces the version that produced the suggestion.
  • Accept, adjust, or reject. Accept a candidate as shown, move its start and end, relabel it, or reject it with a reason.
  • Add what was missed. Mark an event the model did not find, so the record reflects what actually happened.
  • Flag unusable footage. Mark a stretch as covered, out of frame, or too dark so it is kept out of the data.
  • Review from the keyboard. Step through candidates quickly, and build a highlight reel of the clips worth sharing with the team.
  • BCBA for high-consequence labels. Labels such as contact with a person, contact with the child's own head, or physical guidance need a BCBA to review.
  • Session data, then evidence. Accepted events are added to session data while the session is open. After the visit, a BCBA can add one as a labelled late entry until the note is submitted for approval and before a claim is submitted. After that, it stays as confirmed video evidence cited in the BCBA's note review. Nothing unreviewed enters a note, a claim, or a parent view.
  • Show the evidence behind a sentence. On a draft or returned note, select a quote and link a confirmed event. The signed revision keeps that evidence binding. Editing the narrative requires the reviewer to link the revised wording again.
  • Resume a saved review. If adding a confirmed event to session data is interrupted, an authorized reviewer can resume it after the source is verified. The original decision and reviewer remain recorded, and retries do not add the event twice.

Why the three layers stay apart.

A candidate carries three separate things. Keeping them apart means a reviewer can always see what was seen before deciding whether to trust the label. The BCBA note review cites only accepted events, and trends use clinician-verified events only.

  • Observation. What is visible in the clip, in plain words. Example: an adult places cards on the table.
  • Model label. The label the model suggests for that moment. It is unverified and is shown as a suggestion.
  • Clinician interpretation. What it means for this client. Only a clinician writes this, and the model never does.

It reports what is visible, and nothing more.

What it will never do:

  • Infer a diagnosis or severity.
  • Infer why a behavior happened, or what anyone meant by it.
  • Read emotion from a face, or pain from how someone looks.
  • Judge progress, or the quality of a caregiver or therapist.
  • Raise a real-time alert. It reviews recordings after the session, and it is not a safety monitor.
  • Monitor staff. It looks for the client's targets, not for people.

Consent comes first, and the raw video does not stay.

  • Three separate consents: recording, AI analysis, and detector improvement. One does not imply the others.
  • Video only. No audio is processed or sent anywhere.
  • Raw video is deleted after the clinic's retention window, 30 days by default, or when recording consent is withdrawn. Evidence clips of confirmed events stay with the clinical record.
  • Clips open through signed, short-lived links, and every access is audited.
  • Analysis goes directly to the model provider with no stored responses.

A worked example, one candidate from clip to evidence.

An example with invented figures, to show how the chain works. It is not a customer result and not a prediction for your clinic.

  1. Evidence

    A consented recording of a session. A client's active target is requesting an item. The footage passes the quality scan.

  2. Rule

    A candidate appears at 06:42 in the recording. Plain code confirms the label exists, the target belongs to this client and the times fit the recording.

  3. Proposal

    Observation: a child hands a card to an adult and the adult gives a cup. Model label: a request, unverified. A short evidence clip is cut.

  4. Approval

    The BCBA watches the clip beside the definition, moves the start by a second and accepts it. Only the BCBA writes what it means for this client.

  5. Action

    The accepted event is added to session data while the session is open, or as a labelled late entry later, until the note is submitted for approval.

  6. Audit

    The recording hash, model ids, prompt version and definition version stay with the candidate, and every access to the clip is audited.

Availability

Questions about session video.

What does ABACare Vision do with session video?
After a session, it reviews a separately uploaded, consented recording and proposes timestamped candidate moments for the client's active targets, each with a short evidence clip. Telehealth calls are not recorded. A clinician accepts, adjusts, or rejects each one. Accepted events are added to session data while the session is open. After the visit, a BCBA can add one as a labelled late entry until the note is submitted for approval and before a claim is submitted. After that, it stays as confirmed video evidence cited in the BCBA's note review.
What consent does session video need?
Three separate consents: recording, AI analysis, and detector improvement. Analysis does not start without the first two, and detector improvement is its own choice.
How long is raw session video kept?
Raw video is deleted after the clinic's retention window, 30 days by default, or when recording consent is withdrawn. Evidence clips of confirmed events stay with the clinical record. No audio is processed.
What will it never infer?
It does not infer diagnosis, severity, behavioral function, intent, emotion from faces, pain, progress, or the quality of a caregiver or therapist. It is not a real-time alert or safety monitor, and it is not staff surveillance.

Talk through it with your clinical team.

We will walk through consent, retention and the review studio on synthetic data, and what needs to be in place first.