Platform
Software that runs hospital care at home.
Most hospital-at-home tooling was built for academic centers with a dedicated team of twelve. This was built for a program with two nurses, a director who also runs case management, and a CFO who needs the numbers by Thursday — whether that program starts next quarter or started two years ago.
Find every eligible patient
Your ED and inpatient census move faster than any human can screen, so eligible patients go upstairs to a bed before anyone asks the question. The model reads both continuously and surfaces the patients who qualify for care at home, with the evidence that led to each call. Your clinicians confirm or decline in one click.
- Continuous AI screening across ED, observation, and inpatient census
- Every candidate arrives with its reasoning: diagnosis, acuity, distance, support at home
- Declines are recorded, so the criteria converge on what your clinicians actually accept
Moves: candidates surfaced per day — the top of the funnel every stalled program is missing.
Community-acquired pneumonia
Meets criteria: CURB-65 of 1, room air sat 95%, lives with spouse at home, within service radius.
Cellulitis, lower extremity
Meets criteria: afebrile 18h, IV antibiotics only, no ICU history. Awaiting social work note.
CHF exacerbation
Borderline: diuresing well, but lives alone — needs caregiver confirmation.
COPD exacerbation
Meets criteria: off BiPAP 24h, ambulatory, home O₂ already in place.
Run the day safely
A program at census generates hundreds of signals a day: vitals, device alarms, missed visits, patient messages. They arrive triaged into one inbox, ranked by what could actually go wrong. Every alert your team dismisses teaches the model to suppress the next one like it, so the noise falls as the program runs.
- Severity tiers, each with a documented escalation path and an owner
- False-alarm suppression trained on what your nurses actually dismiss
- Visit routing that accounts for drive time and traffic, not straight-line distance
Moves: false-alarm rate — the number that decides whether your nurses trust the system by week three.
SpO₂ 88% sustained 6 min, no response to call
2 min ago · RN Alvarez notified
Missed evening visit — no answer at door
14 min ago · assigned to on-call
HR trending up 12 bpm over 4h, afebrile
38 min ago · queued for rounds
Patient message: question about antibiotic timing
1h ago
Cuff disconnected — reseated by caregiver, confirmed
2h ago · closed by RN Okafor
Grow the census
Programs stall between eight and twelve patients, and the leak is almost always referrals. Every referring clinician sees what happened to the patient they sent, which is the only thing that makes them send a second one. Enrollment is built on research into what actually moves a family from no to yes — not on a consent form handed over at discharge.
- Closed-loop referral status returned to the referring service automatically
- Consent and enrollment a family can complete at the bedside in minutes
- A daily census forecast, so staffing matches next week instead of last month
Moves: average daily census — the only number that moves break-even.
Morning vitals + IV
IV antibiotics
Labs + assessment
CHF check + weights
Telehealth rounds
Evening vitals
Prove the value
Your CFO has about eight questions, and none of them are about software. They get answered from the same data the program runs on, so the operational view and the board view can't drift apart. Exportable, current as of this morning, and defensible in a finance committee.
- Contribution margin per episode, measured against the inpatient alternative
- Length of stay, escalation rate, readmissions, and patient experience in one view
- Payer mix and waiver compliance reporting as a single export
Moves: contribution margin per episode — the number that keeps the program funded.
Average daily census
11.4
+2.1 vs last month
Contribution margin / episode
$3,180
+$240
Average length of stay
4.1 d
−0.3 d
Escalation to inpatient
6.2%
−1.4 pts
30-day readmission
9.1%
−2.0 pts
Patient experience
4.8 / 5
n = 212
Referral conversion
63%
+9 pts
Days to break-even
Month 7
1 ahead of plan
Under the hood
What's actually running.
Four properties that decide whether clinical AI gets used or quietly turned off in week three.
Continuous ingestion
HL7v2 and FHIR feeds from your EHR, normalized into one patient timeline. Streaming, not a nightly batch — a patient who becomes eligible at 2pm surfaces at 2pm, not tomorrow morning after the bed is gone.
Evidence-first inference
Every model output ships with the features that produced it. A clinician sees "CURB-65 of 1, room air sat 95%, spouse at home" rather than a score with no provenance. That is a deliberate non-device clinical decision support posture, and it is also the only way clinicians ever come to trust a model.
Feedback as training data
Accepts, declines, dismissals, escalations, and outcomes are all captured as labels by the same system that runs the program. Most clinical AI is trained once and deployed blind. This is trained by the people using it, every day they use it.
Federated improvement
Model updates travel across the network. Patient records do not. Your tenant stays your tenant, and the platform still gets better every time another hospital joins.
Trust
Your data stays yours.
Tenant isolation
Your patient data lives in your tenant. It is not pooled with other hospitals, and it is not used to train models that serve anyone else's patients. What improves across the network is learned from de-identified aggregates and model updates, never from raw records leaving your walls — and any hospital can opt out of contributing without losing access to the platform.
BAAs, signed before anything connects
We sign a Business Associate Agreement with every hospital before a single interface is turned on. Access is role-scoped, least-privilege by default, and reviewable by your compliance team at any time.
Clinicians always decide
Nothing here diagnoses, and nothing here treats. The AI surfaces candidates and ranks signals; a licensed clinician at your hospital makes every clinical decision. Every recommendation shows the data underneath it, so a clinician can disagree with the reasoning rather than with a black box.
Audit trails you can actually read
Every view, every alert action, every eligibility decision is logged with a user and a timestamp, and exported on request. If a regulator or your own quality committee asks what happened on a given night, the answer is one query away.
doorstep.png
A clinician greeted warmly at a front door in golden-hour light