Technology at Roost

The care behind the care.

Good technology gives people better information and then gets out of the way. Roost keeps the details in one place, so a family and a caregiver can focus on each other.

Three systems, one platform

  1. 01

    Getting to know each caregiver

    Affinity AI

  2. 02

    Who you see, and why

    Temporal AI

  3. 03

    What the platform watches

    Horizon AI

More understanding.Human judgment, at every step.

Three systems support one relationship: the caregiver, and the family who chose them.

01 / Affinity AI

Getting to know each caregiver

Before a profile can be listed, Roost walks a caregiver through the questions, their documents and a criminal background check, in one path with nothing typed twice. Affinity AI organises what comes back so the person reading it has the whole picture in front of them. The reading, and the decision, are theirs.

  • One path, nothing typed twice
  • Documents and checks kept in view
  • Every application read by a person
Meet Roost as a caregiver →
02 / Temporal AI

Who you see, and why

Availability is only part of the picture. Temporal AI puts the caregivers who are free near the care address in front of you with what each of them told us, so you can start the conversation you would have started anyway. It recommends nothing, decides nothing, and never sees a protected characteristic. You choose who to message.

  • What you said the person needs
  • What each caregiver said about their work
  • Distance from the care address
Find a caregiver →
03 / Horizon AI

What the platform watches

Every alert carries the readings it was raised from and a note saying it names no condition and directs no treatment. One reading is never a baseline. What a family sees is their own visits and what they were told, in plain words, not a score.

  • Visit records in one place
  • Every flag carries its readings
  • Not a diagnostic tool
Find a caregiver →

The everyday details. Connected underneath.

Underneath the list sits one record of each booking: what the family and the caregiver agreed, the hours the visit ran, and the charge that followed.

  • Hours settled before a charge

    A visit is charged only after the caregiver files a note and the family confirms the hours, or lets the window to question them close.

  • A note that stays as written

    The caregiver's note after a visit is never edited. A correction is added beside it, so the first version stays readable.

  • Paid to the caregiver

    The family pays the caregiver's own account. Roost's platform fee is on the screen before they book, and the receipt is itemised.

Explore the platform in detail · interactive demo

For a closer look at the operations: follow an illustrative caregiver, visit, and recovered dollar. Agency names, locations, and figures in this demo are examples, not a list of Roost service locations or a promise of results.

Interactive demo

One caregiver. One visit. One dollar.

Roost owns US home care agencies and runs them on one operating system. Follow one caregiver, one visit, and one dollar through the machine — and watch every step land on the same append-only event spine.

0/18 events on the spine

or jump to any stage — every step lands on the spine below.

The estate

Ownership + tenancy

Roost buys home care agencies and runs every one of them on the same platform. Each agency's rows are walled off by row-level security — scope comes from the signed-in session, never from a request. The platform runs on synthetic data and holds no real health record, so no health-privacy obligation has been triggered; the controls those obligations ask for are built in now rather than retrofitted later.

Blue Heron Home Care
NY · Roost's own visit verification
Casa Verde In-Home
TX · state-mandated aggregator
Lakeside Caregivers
acquired — onboarding gate
cross-agency read → blocked by row-level security
One platform — Roost OS · one append-only event spine

Scope from the session, never the request. One covered entity, one compliance program — a compromise is bounded to one agency's rows.

writes to the spine — the estate's append-only event log:agency.admitted

A caregiver joins

Affinity AI

M. Alvarez applies, and walks the same gated path as every caregiver. The FCRA state machine is enforced at the database boundary: no disclosure and authorization, no background check; no human adjudication, no adverse action. Steps taken out of order are refused, not warned about.

✓
Application
AI drafts + structures
✓
Documents
licenses, I-9
✓
Disclosure
standalone — nothing else in the document
✓
Authorization
signed before any order
✓
Background check
status + link only; the report stays with the provider
✓
Human adjudication
a person decides — always
✓
Approved
ready to match
✕ order a check before authorization? Refused — the state machine has no such transition
Pre-adverse notice + report copy + rights summaryFinal adverse — evidence chain re-derived

The wait is recomputed from the stored date. No stored value can shorten the legal floor.

writes to the spine:disclosure.signedauthorization.signedcheck.orderedadjudication.recorded

Choosing who comes

People Predictor

The Predictor lists the caregivers who fit what the family asked for, with the reasons beside each one. A person makes the choice, never the software, and the record keeps what was shown, who was chosen, and how the booking went.

M. Alvarez
92
J. Chen
84
R. Okafor
77
dementia experience3.2 mi awayschedule fits Tue/Thulanguage: Spanish

Versioned and logged. Weights are stamped on every recommendation; followed-vs-overridden outcomes build the evidence view — empty until real data exists.

A person decides

Follow the recommendation, or override it — an override must carry a reason (the database refuses one without it).

recommended: M. Alvarezhuman chose: M. Alvarezoutcome at 90 days: pending
writes to the spine:match.recommendeddecision.recorded

A visit happens

Roost OS · scheduling + EVV

Alvarez's Tuesday visit is scheduled, then electronically verified (EVV): clock-in, clock-out, location. When a correction is needed, the original record is never touched — the edit is appended alongside it, so the trail always shows what was recorded first and who changed what.

Scheduled
Verified (EVV)
clock-in 9:02 · GPS ✓clock-out 10:04 · 62 min
Correction? Appended, never overwritten:evv_edits + row · reason: "caregiver forgot to clock out"original row: still there
writes to the spine:visit.clock_invisit.clock_outevv.edit_appended

The visit becomes a claim

Roost OS · claims

Fourteen pre-submission checks scrub every line; a blocking finding makes submission impossible until it is fixed. Minutes are summed across every backing visit, then rounded by the payer's own policy — never a global rule. And a claim is not “filed” until the payer's claim-status response — the 277CA — says A2.

Scrub — 5 of the 14 pre-submission checks

  • Prior authorization on file
  • Units within remaining authorizationBLOCKFIXED
  • EVV minutes back every billed unit
  • Caregiver credential current on date of service
  • Within payer's timely-filing window

A block makes submission impossible. Here the auth is topped up and the line re-scrubbed — blocks that would have become denials are the ledger's "claim revenue protected."

The same 62 verified minutes — two payer policies, two claim formats

NY personal care 837I · institutional

Nearest hour, 1-unit floor under 30 min — sum across all backing visits first

62 min → 1 unit (hour)
TX Medicaid 837P · professional

Quarter-hour units

62 min → 4 × 15-min units

Then the payer answers — and only one answer means "filed"

999
syntax accepted
not filed
277CA · A1
receipt only
not filed
277CA · A2
accepted for adjudication
FILED
835
remit — $33.20 paid
paid
writes to the spine:claim.scrubbedclaim.submittedack.999ack.277ca_a2remit.posted

Dollars come back — and fund the raise

Recovery Ledger + Wage Loop

Every recovered dollar is attributed to the mechanism that recovered it and traceable to a source event — an unattributable dollar is not counted. A share of recovered dollars funds a targeted wage increase, and the spine measures the retention response against a matched control.

1 · Recover One agency, one year (demo figures)

Turnover cost avoided$126,000
Claim revenue protected$84,000
Denials recovered$41,000
Missed visits prevented$38,000
Leakage closed$29,000
Recovered$318,000

Auditable down to the event. Every entry links to its events and, where a human acted, to the decision row — an unattributable dollar is not counted.

2 · The raise it funds Median turnover by wage band

$15.00–17.99/hr
84.4%
$18.00–19.99/hr
53.0%

A $1–2/hr raise crosses a ~31-point cliff. The loop funds it — and measures the response against matched controls.

3 · Measured, loop closed

Recovered $+$1.50/hr cohortRetention measuredLedger turnoverline — loop closed
writes to the spine:ledger.entrywage.increaseretention.measured

One append-only event spine

Everything above wrote to the same spine — 0 of 18 events lit. Tap a dim event to jump to the stage that writes it.

When a human decides, the spine records it — recommendation, choice, outcome. That is what makes the Recovery Ledger auditable down to the event, and the Wage Loop a measurement instead of a promise.

Technology should bring care closer.

Start with the people. Tell us who needs company and see who is free near them.

Technology — the care behind the care · Roost