Ospia HMS · AI-native hospital management
Ospia is the hospital management system built AI-native from the first line of code. Seventeen autonomous agents watch your wards, pharmacy, revenue and compliance around the clock — they draft, your people decide, and every step lands in a tamper-evident audit chain.
no black boxes · every proposal answers “why?” · propose → approve → execute
Product demonstration: Ospia’s agent approval console. An AI agent proposes an action, a named staff member approves it, and it executes with an audit-chain entry.
The gap
Every hospital already has a system of record. Almost none has a system that watches the record and acts on it.
Notes, ICD codes and discharge summaries get written hours after the patient left — from memory, under fatigue, into a form with forty fields.
Reorder levels were set once, years ago. The first person to learn a drug ran out is the patient standing at pharmacy.
Aging invoices, stalled TPA preauths, unsigned summaries blocking record release — discovered in audit, weeks after they could have been fixed.
The agent roster
Ospia ships with seventeen role-based agents — a leadership team in software. Each one watches its beat continuously and files proposals into the same approval queue your administrators already use.
Every agent is propose-only by construction. None of them can touch a domain table. What they can do is notice — at 3 a.m., on a Sunday, during the festival rush — and have a fully-reasoned proposal waiting when your team arrives.
Native AI capabilities
Most vendors added an assistant to a twenty-year-old system. Ospia was designed around the AI — with the guardrails a hospital actually requires.
Speak through the consultation. Walk out with a structured SOAP note, ICD-10 suggestions and a draft prescription — with red-flag symptoms surfaced as you talk. Nothing enters the record without clinician sign-off.
“Which departments had the longest lab turnaround last month?” Ask in English; get a governed answer. Questions compile to read-only SQL behind a five-layer guard and run under the asker’s own permissions — pin any answer as a live dashboard.
A real-time mirror of the hospital — ward heatmaps, OPD flow, theatres, fleet, duty coverage. Rehearse tomorrow before you live it: close a ward, add beds, absorb an OPD surge or a staff absence, and see the impact first.
Drug–drug interactions, allergy cross-reactivity, duplicate therapy, renal dosing and pregnancy checks — at the moment of ordering. Serious alerts block until the prescriber records an override reason, preserved for the medico-legal record.
“Any invoice over ₹50,000 needs CFO sign-off.” One English sentence becomes a validated workflow, approval chain, intake form — or a whole department with wards, beds and services, set up in a single approved transaction.
No-show risk on every booking, bed-demand forecasts, self-tuning reorder levels — each prediction ships with its named factors, so the front desk and the stores team can see exactly why before they act on it.
Photograph the paper registration form; AI transcribes and normalizes it for reception to verify. Aadhaar is reduced to last-4 at parse time and the image is never stored.
Referral letters, old discharge summaries and lab reports are classified and parsed into the chart — diagnoses, medications, test values — for the clinician to confirm.
Snap the BP machine or pulse-ox; readings are extracted, unit-converted and range-checked, then queued for nurse review. No new hardware required.
Qure.ai-class vendors push chest X-ray, ECG and stroke findings over per-vendor HMAC-signed URLs — CDSCO-aligned provenance (model, version, confidence) on every result, and a mandatory clinician review queue in front of the chart.
A FHIR R4 read/search API — Patient, Encounter, Condition, Observation, MedicationRequest — plus an outbound HL7 ADT feed, so your LIS, PACS and public-health reporting connect natively.
HMAC-signed webhooks fire on domain events with a full delivery log, and an OAuth-style client registry issues scoped tokens — integrations get exactly the access you grant, nothing more.
The safety model
Three invariants govern every AI capability in Ospia. They are enforced in the architecture, not the policy manual.
No agent can write to a domain table. Proposals are deduplicated, carry a risk class, and answer “who, why, what, with which inputs” before anyone decides.
Execution requires a staff member holding approval rights. Hospitals may opt in to auto-approval for low-risk actions only — and that policy is itself an approved, audited setting.
Approved actions run through a fixed catalog of typed executors, re-validated against their schema at execution time, inside the caller’s tenancy — and captured by the audit trail.
Every mutation hash-chained; the chain re-verifiable end to end.
Postgres RLS scopes every query to hospital and user context.
AES-256-GCM on personal data, separate keys for audit and PII.
Patients live on a separate identity plane — portal tokens are cryptographically useless on staff APIs.
Physical tenant isolation: your data never shares a database with anyone.
Built for India
Not localized after the fact — Indian statute is in the data model. DPDP, ABDM, NMC and GST behaviors are enforced by the system, and watched by the agents.
Patient onboarding is ABHA-ready, and every use of personal data traces to recorded DPDP consent — in registration, the portal, and telemedicine.
Generic-name prescribing in capitals, prescriber statutory identity on every document, serialized certificates — and the 72-hour record-release clock escalated automatically before it lapses.
HSN/SAC on every line, CGST/SGST vs IGST split by place of supply, e-invoice payloads with deterministic IRN, and a TDS/TCS ledger — statutory accounting, not an export.
Built-in WebRTC video — media never touches the server. Consent is gated before the call, Schedule-X drugs are hard-blocked in-call, and patients join via a WhatsApp code.
Interface in ten Indian languages, appointment and report notices over WhatsApp/SMS, a PII-free waiting-room token board, voice-first booking for low-literacy patients, and prescriptions explained in plain Hindi or English.
Policies, preauthorization and a full claim state machine — with the revenue agent chasing stalled preauths and aging invoices before month-end finds them.
CAPA tracking, a live risk register, HAI surveillance and antimicrobial stewardship — the evidence trail assessors ask for, produced as a by-product of daily work, not a pre-audit scramble.
Attendance and duty rosters, Indian payroll structures, recruitment and medical-staff credentialing — with watchdogs that flag registration renewals before they lapse.
One record, whole hospital
Packaging
Every plan ships all 25+ modules, all 17 agents and the full safety model. Tiers differ in deployment and scale — not in which capabilities you're allowed to have.
No per-video-minute fees, no per-message AI charges. Teleconsults run peer-to-peer, and deterministic fallbacks keep the whole system working even with no AI model configured.
Straight answers
You do, in the plainest sense: each hospital runs on its own physical database, self-hosted if you prefer. Everything is exportable, and the FHIR R4 API means you're never locked in — leaving is as open as arriving.
Write to clinical data on its own. Agents can only file proposals; execution happens through a fixed catalog of typed, re-validated executors after a named human approves. Auto-approval exists only if you opt in, and only for low-risk actions. Nothing the AI drafts enters a patient record without clinician sign-off.
The AI here isn't a feature on the side — the propose-only agents, the whitelisted executors and the hash-chained audit trail are the architecture. A bolt-on assistant can answer questions; it can't watch your wards at 3 a.m. and have a reasoned, auditable proposal waiting at handover.
The hospital keeps running. Ospia is a complete HMS without any external AI: the scribe falls back to a deterministic clinical parser, patient-facing explanations use a template engine, and every core workflow is plain software. AI adds speed; it is never a dependency.
By mechanism, not policy document: generic-name prescribing and prescriber identity are enforced at write time, the NMC 72-hour record-release clock is escalated by an agent, DPDP consent is a ledger consulted before use, Aadhaar is reduced to last-4 at capture, and every mutation lands in a hash-chained audit log you can re-verify end to end.
Yes — we seed a full demo hospital with synthetic clinical data (patients, admissions, pharmacy stock, invoices) and put the agents on duty. Your team explores real workflows with zero real patient data.
Next step
A walkthrough takes forty-five minutes: we seed a full demo hospital, put the agents on duty, and let your team ask Ospia the questions they ask their MIS today. Bring your hardest month-end.
demo seeded with synthetic data · zero real patients · self-hosted or managed cloud · one database per hospital · 237 automated checks on every release