Harvis AI · The intelligence layer

Not 4 AI products.
One intelligence.

Harvis AI is a layer, not a feature. It sits between your live hospital data and the people who have to act on it — reading vitals, labs, orders, beds and invoices as they happen, and putting the one thing that matters in front of the one person who can do something about it.

15+ models in production. No prompt, no query, no chat.

Harvis HIS is the record. Yakko drafts into it, Gooseberry re-reads it, BucketAI ranks it, Harvis Me carries it home.

Book a 20-minute demo Follow the morning ↓
08:41 · Murugan arrives 1Arrives 2Consult 3Second read 4Discharge → home 5At home 6Next visit · trend on the chart
Inside the hospital · one record
Yakkodrafts into it Gooseberryre-reads it BucketAIranks it
Harvis HIS · the recordHRV-024815
Murugan R.58 / M · T2DM, hypertension⚠ Penicillin
HIS · 08:41Token A-42 · BP 138/86 · HbA1c 8.4%
Yakko · 09:02Clinical note · 5 sections · signed
Gooseberry · 09:14K⁺ 6.2 on spironolactone · cited · task C-2291
BucketAI · +8 wksScore 0.88 · cardiology consult Tue 14:30
Harvis Me · from home · by consentRHR ▲ 9 bpm · adherence 94%
At home · Harvis Me
Harvis Me · Muruganhome
From the hospitalDischarge plan · follow-up Tue 14:30Prescriptions on the phone, not a folded printout
Resting heart rate · your band vs you ▲ 9 bpm · since 14 Jul
Adherence · 148 days94%Back on the chart · by consent
  1. 1Arrives · 08:41Harvis HIS — the chart is already full.
  2. 2Consult · 09:02Yakko drafts the note; the doctor signs.
  3. 3Second read · 09:14Gooseberry cites the finding, opens task C-2291.
  4. 4Discharge · the doorPlan, prescriptions and follow-up date go to the phone.
  5. 5At homeHarvis Me learns the baseline and dates the drift.
  6. 6Next visit · +8 weeksBucketAI booked it; the home trend is on the chart.

The care does not stop at the door. One record carries the patient in, home, and back.

Interface vignettes show illustrative data, not real patients or records.

15+Models in production
3M+Patient records the models work from
75+Hospitals live
55Modules in the record underneath

08:41 · Front desk → doctor’s desk

The room already knows him.

One UHID — the single hospital ID a patient keeps for life — pulls six years forward. Nothing is re-asked at the counter, and the doctor’s screen is full before the door opens — including five months the hospital never saw.

Front desk wrote it at 08:41. Dr. Priya reads it at 09:02. The pharmacist will read the same row at 09:26. Every line on the chart names the product and the person that wrote it — which is what “one record” means when you can see it.

The last three rows are the part a hospital cannot get any other way: Harvis Me has been watching the patient’s own baseline at home, and put the drift on the chart — by consent — before anyone asked. A consultation is twenty minutes and a memory test. This one starts from evidence.

  • Created once. Never re-asked. The UHID, the allergy and the payer were written at the counter and read by everyone after.
  • Pre-auth confirmed before the consult, not after — the hospital side of the same row.
  • Two rows still to come this morning: the note Yakko will draft, and the second read that runs the moment it exists.

How Harvis HIS keeps that row true → What Harvis Me adds to it →

Harvis HIS · OPD chart · Dr. Priya09:02
Murugan R.58 / M · UHID HRV-024815 · known T2DM, hypertension⚠ Penicillin allergy
HIS · front desk · 08:41Token A-42 · General medicine · Star Health · pre-auth ok
HIS · vitals · 08:52BP 138/86 · HR 82 · SpO₂ 97%
HIS · lab · prior visitHbA1c · last 8.4%
HIS · pharmacy recordActive scripts 4 · incl. spironolactone 25 mg OD
HIS · allergy register⚠ Penicillin
Harvis Me · by consent · 5 monthsRHR ▲ 9 bpm since 14 Jul · exertional episodes logged from 9 Jul · 6 events
Harvis Me · by consentAdherence 148 days · 94% · missed doses dated
Yakko · 09:02Clinical note — drafted after the consult, signed by Dr. Priya
Gooseberry · 09:14Second read — the moment the note exists
7rows on screen before the door opens0re-asked at the counter2still to come this morning

Interface vignettes show illustrative data, not real patients or records.

09:02 → 09:26 · Consultation

The doctor talks. Nobody types.

Forty seconds after the patient stands up, there is something to sign. Twelve minutes later, when today’s potassium result lands, the chart has been read back against that note, a finding cited, and a named clinician has a task with a due time.

What ran in those twelve minutes

The moment a note exists, the same six moves run against the chart behind it — vitals, labs, history, allergies, prescriptions, pathways — in a fixed order, every time. Gooseberry is a deterministic second read from this patient’s chart data — not a diagnosis, and not a replacement for your judgment. Every claim quotes the row it came from; a claim with zero citations cannot even be stored.

Gooseberry proposes; a named clinician disposes — accept, reject or override, and the last two need a written reason. It writes nothing to the chart, ever. The escalation is not a banner: it is clinical task C-2291, owned by Dr. Priya, due in thirty minutes.

Gooseberry supports clinical decision-making. It does not replace physician judgment, hospital protocol, or regulatory review. All suggestions remain subject to clinician sign-off.

Watch a full run, move by move → How Yakko drafts the note →

PrescriptionFinding · urgent

Potassium 6.2 mmol/L while on a potassium-sparing diuretic

Reasoning. Today’s panel returned K⁺ 6.2 mmol/L (reference 3.5–5.1). The active prescription list includes spironolactone 25 mg OD, started 11 days ago. Both facts are quoted below.

LAB_RESULT“6.2 mmol/L”lab_results · 09:14 today
PRESCRIPTION_ITEM“Spironolactone 25 mg OD”prescriptions · 21 Aug

Coverage. Screened against 1,284 interaction rows currently loaded for this tenant. Renal function was not read by this check.

AcceptRejectOverrideReason required to reject or override
EscalatedClinical task C-2291 · Dr. Priya · due in 30 min · Accepted 09:21

Interface vignettes show illustrative data, not real patients or records.

How they work together

Five passes over one record.

Not a shelf of products you buy separately and wire together — five passes over the same record, each picking up exactly where the last one stopped. Here is what each one hands to the next.

Murugan R.58 / M · UHID HRV-024815 · known T2DM, hypertension
Illustrative
Harvis HIS08:41 · The room knows him

Front desk writes the token and the confirmed payer; nursing adds today’s vitals at 08:52.

Hands forward: UHID · vitals · allergyWritten once, read by everyone after Yakko09:02 · The note writes itself

Drafts five coded sections from the conversation; Dr. Priya edits and signs.

Hands forward: signed note · 5 sections40 s after the patient stands up Gooseberry09:14 · The chart is read back

Six moves over the whole chart; the finding quotes its rows and escalates to a named owner.

Hands forward: cited finding · task C-2291Nothing written to the chart BucketAI+8 weeks · Not forgotten

Scores the whole panel; the follow-up is booked into the same calendar.

Hands forward: cohort score · a booked follow-upA reason the receptionist can read aloud Harvis MeOngoing · The record goes home

Discharge instructions and the follow-up date on his phone. His home trend was already on the chart this morning — and keeps accruing for the next visit.

Hands forward: home trend · biomarkersBy consent, reversible at any time

Five products. One row in the database. No integration project, no nightly sync, no export-and-reimport — every pass above wrote to the same live record. Eleven things gained, zero re-typed.

Every tierCore AI — risk scores, reorder alerts, leakage detection — is on every plan. Professional +Yakko, Gooseberry and BucketAI start at Professional; custom model training is Enterprise. DirectHarvis Me is bought by individuals — early access. See pricing →

A model bolted onto someone else’s HIS sees a nightly export. The Harvis layer reads the row as it is written — which is why the passes above can share one record at all. Walk the same patient through the departments →

Interface vignettes show illustrative data, not real patients or records.

One row, five passes

What each engine reads — and what it is allowed to write.

“One record” is checkable. This is the truth table of the platform: which record objects each pass reads, which it writes, and how — as a draft for signature, a task with an owner, a booking, or with the patient’s consent.

Gooseberry has no write path to the chart. It records a decision and, when it escalates, opens a real clinical task — never a prescription, a diagnosis or an order.

No note enters the record unsigned. Yakko’s only write is a draft, and the draft becomes the record when a doctor signs it — 0 unsigned entries.

Core models write in-flow. TIMI, GRACE, CHA₂DS₂-VASc, NIHSS and PHQ-9 computed from the chart as it is written; reorder alerts and unbilled flags raised against the row, not emailed about it.

11:40 → 18:00 · The rest of the shift

The consult is over. The record is not.

Three more desks read the same row that afternoon — none of them opened a chat window.

11:40
Core AI · central store Enoxaparin will cross its reorder point on Saturday

The purchase order is raised on Tuesday, while there is still a week of cover — instead of on Saturday, at emergency rates. The reorder alert fires before the stock-out, not in the post-mortem — with cold-chain and expiry safeguards on top, and drug-licence expiry alerted at 90 · 60 · 30 days.

16:20
Core AI · billing 184 delivered services missing from open invoices

Leakage detection catches the service that was delivered but never billed — before the invoice closes, so it is a correction rather than a credit note and a difficult phone call. Denial analytics reports rejection rate by denial reason and by TPA, so the same mistake is not re-filed.

18:00
BucketAI · front desk Tomorrow’s call list is already built

The whole panel scored into 27 buckets — not a sample. Twelve high-risk cardiac patients with no visit in 60+ days become a call list, each with a plain-language reason the receptionist can say out loud. Eight weeks on, Murugan is on it: score 0.88, cardiovascular, and a cardiology consult booked into the same calendar for Tuesday 14:30. A bucket is not a diagnosis — it is a reason to make contact.

Interface vignettes show illustrative data, not real patients or records. Cohort sizes and values shown are examples, not results from a named hospital.

Value, by desk

Three balance sheets. One record.

The same row pays out differently depending on who is reading it. None of the payouts require anyone to ask.

🩺 The doctor🏥 The hospital📱 The patient
Harvis HISHistory and allergy already on screenOne UHID pulls six years forward.55 modules · 770+ screensStatutory registers generated on their own cadence.Created once. Never re-asked.The pharmacist sees the same potassium the doctor saw.
Yakko~2 hrs of typing returned, daily3.2 hrs → 40 min per doctor per day.~11,000 hrs a yearAcross a 20-doctor department.Eye contact instead of keyboardsIn தமிழ் and English.
GooseberrySix moves on every caseEvery claim cited · 0 written to the chart · a clinician can overrule any of it.A decision ledgerRun, citation, decision, reason, name — for the quality committee, months later.Owned within twelve minutesA dangerous potassium result had a named doctor on it, with a 30-minute deadline.
BucketAIEvery score with its reasonsWeighted chart factors, in plain sight. Overrule in one click.The whole panel, ranked by need27 buckets · a worklist that books into the same calendar.Not forgottenA booked follow-up instead of a lost one — called, with a reason you can understand.
Harvis MeFive months of trend on the chartBefore the patient sits down — by consent.Continuity in both directionsDischarge and follow-up on the phone; home panels back on the chart.2 of your 140 blood markers need attentionThe other 138 are fine — and it says so.

Clinical decision support is the one row we refuse to price. For the hospital ledger, the site’s published averages apply — ₹4.2L revenue recovered per month (payback is typically reached in about six months). Typical movement reported by hospitals in the first quarter after go-live. Your numbers depend on your case mix and starting point — we size them with you before you sign anything.

Interface vignettes show illustrative data, not real patients or records.

How we keep it honest

A model that can’t be interrogated is a liability.

Any model degrades as your population drifts. The only question that matters is whether it admits it. Three gates stand between a signal and the chart — and the fleet reports its misses, per model.

In · the live row
VitalsLab resultsPrescriptionsBedsOrdersInvoicesStockClaims
Show the source Passes: a claim with at least one verbatim citation — table, record id, value, clinical timestamp — from one of eighteen record classes. Stops: a claim with zero citations. It cannot be stored — a database constraint, not a review step.
0uncited claims
A human signs Passes: Yakko’s draft for signature; Gooseberry’s suggestion for a decision; BucketAI’s worklist a clinician can overrule in one click. Stops: anything auto-applied. 0 unsigned entries. 0 written to the chart by the second read. Four decision states, two of them requiring a written reason.
0auto-applied
State the limits Passes: a card that says what it screened — “1,284 interaction rows; renal function was not read.” Insufficient data names the missing fields and is not decidable. Stops: “nothing flagged” masquerading as “checked and clear”. A run made entirely of don’t-knows is flagged, never clear. Confirm rate is reported per model — not one flattering average.
0“clear” without stated coverage
Out · the chart

No clinical fact — a note, a diagnosis, an order, a prescription — is written by the layer. Those are written by a named human: Dr. Priya, the pharmacist, the front desk. The layer writes scores, alerts, tasks and bookings, all shown in the ledger above.

Five of the eight BucketAI specialty models · confirm rate, last 30 days · illustrative8 models · 1 drifting
Cardio riskcardio-l2 · v3.796.1%
Diabetesdm-router · v3.4 · drifting · queued for retraining93.4%
Onco screeningonc-scr · v2.988.2%
Mental healthphq-net · v1.884.0%
Lifestylelifestyle-b · v2.281.3%
1 driftingThe diabetes model has moved away from its training distribution and is queued for retraining. You are told before its recommendations quietly rot.
84 casesQueued for a clinician to label, because the model was genuinely unsure. Unsure cases go to a human, not to a call list.Highest-value training data you own

Gooseberry is not in this panel: a deterministic second read does not retrain, it is versioned. Its ledger is the audit trail — every run, every citation, every decision with its reason.

Book a 20-minute demo

See your hospital run on Harvis.

No slide decks. We’ll walk you through a live OPD consult, a discharge and a billing cycle — on your specialty, in your language. Here is what typically moves in the first 90 days.

Average OPD wait48 min12 min−75%
Documentation per doctor3.2 hrs40 min2 hrs given back, daily
Manual errors120 / mo~5 / mo−96%
Patient rating3.6★4.6★+1.0

Typical movement reported by hospitals in the first quarter after go-live. Your numbers depend on your case mix and starting point — we size them with you before you sign anything.

Book a demo See pricing

Just want the app? Join Harvis Me early access → Run a hospital on it — or just run yourself on it.

Figures are averages reported across live deployments, not a guarantee. We model your own baseline during the demo. Interface vignettes show illustrative data, not real patients or records.