A Clinical Intelligence Platform Running Entirely on a Hospital's Own GPUs

A patient with chest tightness books pulmonology. The cause is hypertension, so they need cardiology. They use a pulmonology slot, learn nothing useful, and re-book — while the cardiology slot they actually needed was taken by somebody else who guessed differently.

Multiply that across 43 specialties and more than 133,000 outpatient visits a year and it stops being an inconvenience. It is a capacity problem that looks like a demand problem, and no amount of adding clinics fixes it.

The Challenge

The outpatient department of a large multi-specialty hospital network was under real strain. Growing footfall had exposed structural problems that manual processes could not resolve:

• Patients self-selected their specialist from a departmental directory, based on nothing more than which name sounded closest to their symptom.
• No structured pre-screening existed at the point of booking. There was no triage, no intake assessment, nothing between a patient's guess and a consultant's diary.
• Patients arrived without their records — no prior imaging, no previous prescriptions — so consultations were spent reconstructing history that should have been gathered upstream.
• The high-volume departments — General Medicine, Orthopaedics, Neurology, General Surgery — were spending a disproportionate share of every slot on that reconstruction.
• Front-desk queues spilled into adjacent waiting areas.

And the compounding effect: each misrouted patient consumed a specialist slot, the correct specialist then had no capacity, and the total care timeline extended by days.

The constraint that shaped everything. As an accredited institution handling longitudinal patient records — diagnostic imaging, prescription histories — the hospital could not route any patient data to external cloud APIs. Not as a preference. Any solution had to run entirely inside their own infrastructure.

That single requirement rules out most of what is currently sold as clinical AI.

What We Built

Three phases across roughly twenty months, each independently deployable, each adding capability without disrupting workflows already running. The architecture was on-premise from day one rather than adapted to be later.

Phase 1 — Months 1–4: Structured conversational intake

A conversational intake agent inside the hospital's existing patient portal and WhatsApp channel, handling booking for straightforward cases: a keyword classifier over a curated clinical taxonomy mapping common complaint phrasing to departments, a confirmation with pre-visit instructions by SMS and WhatsApp, and a reminder to bring existing reports.

It took routine booking off the front desk immediately. It also demonstrated its own ceiling — a taxonomy still sent a patient to orthopaedics when their presentation warranted neurology, because matching words is not the same as understanding a complaint. That failure is what motivated Phase 2.

Phase 2 — Months 5–20: Agentic diagnostic intelligence

The defining change: from a rule-based assistant to something doing genuine clinical reasoning.

An adaptive symptom interview, asking follow-up questions based on what the patient actually said rather than running a fixed questionnaire. Someone describing lower back pain is taken through duration, radiation pattern and prior history — the questions a clinician would ask, in the order they would ask them.

Differential hypothesis generation. Once there is enough signal, the system maps the symptom cluster to plausible conditions with supporting reasoning — presented to the patient in plain language, and never as a diagnosis.

Specialist matching with the patient still choosing. Two or three profiles from the live roster, with sub-specialisation, experience, available slots, and a short explanation of why this consultant is relevant to your presentation. The patient retains complete autonomy over the choice. The system explains; it does not decide.

A personalised preparation checklist per clinical pathway — bring your OP card, retrieve the MRI from wherever it was taken, bring your monitoring records and current prescriptions. This is the unglamorous half and it removed a great deal of wasted consultation time.

Pre-arrival vitals. On the morning of the appointment, a structured check-in asks the patient to self-report weight, blood pressure if they monitor at home, and anything new since booking — prepopulated into the physician's view before the patient walks in.

Available in English and Malayalam, through both the web portal and WhatsApp.

Phase 3 — In trial: predictive EMR intelligence

Currently a controlled trial in two departments, Nephrology and Internal Medicine, with rollout to Oncology and Cardiology pending outcomes.

As records migrate from paper and semi-structured formats into a structured repository, a Graph-RAG engine builds a temporal knowledge graph of each patient — so a physician can ask "what was the last HbA1c and when?" and get a structured answer without leaving the consultation screen.

The brief is delivered sixty minutes before each appointment, so inference is complete before the consultation window opens rather than competing with live traffic at peak.

Two things it caught during the trial, both worth stating precisely because they show the shape of the value:

• An early renal mass signal. Pattern analysis across serial ultrasound measurements flagged an uncharacteristic growth trajectory in a patient attending for something else. The physician ordered a contrast CT, which confirmed early-stage renal cell carcinoma. It was not the stated reason for the visit.
• A medication interaction. A diabetic patient's newly prescribed antibiotic, dispensed by an external pharmacy, was cross-referenced against their existing metformin regimen. The interaction had not been surfaced by the dispensing system. It appeared in the pre-consultation brief and prompted a dosage review.

What it deliberately does not do: recommend treatment, generate prescriptions, or communicate findings to patients. Everything is labelled as AI-generated and presented for the physician's own judgement. Ingestion is one-directional — nothing generated ever writes back into the authoritative record.

The Infrastructure Problem, and How It Was Solved

Running a full clinical AI stack on-premise means confronting something cloud deployments never have to: the workloads span an enormous complexity range. A slot-availability query and a longitudinal differential are not the same computation, and sizing hardware for the second to serve the first is how on-premise AI becomes unaffordable.

The answer was a three-tier router that classifies each request by required reasoning depth and context size, in under 50 milliseconds, invisibly:

• Tier 1 — quantized models on existing CPU hardware. Booking confirmations, slot queries, FAQ responses, simple extraction. No GPU, no new hardware.
• Tier 2 — GPU. The adaptive interview, differential generation and Graph-RAG synthesis, with embedding workloads batch-scheduled and priority-queued so background indexing never competes with a patient waiting for a reply.
• Tier 3 — frontier API, non-PHI only. A small number of high-complexity tasks containing zero patient-identifiable information. This is enforced architecturally rather than by policy: a PHI-detection filter runs before any Tier 3 call, and anything failing it is handled locally regardless of the quality cost.

That last detail is the one worth copying. "We will not send patient data outside" is a promise. A filter that runs before the call, and that prefers a worse answer to a leaked one, is a control.

The Result

Across the digitally-engaged cohort, measured against the pre-existing intake pathway:

• 59% fewer wrong-specialist appointments
• ~12% shorter average consultation — a conservative average across departments, with high-adoption departments reporting more
• 96.2% intake completion rate
• 3.8 minutes average pre-screening duration
• 60 minutes of lead time on the physician's pre-consultation brief

The 59% is the number that pays for the project, because misrouting was never a single wasted appointment. It consumed a specialist slot, left the correct specialist without capacity, and pushed the whole care timeline out by days.

Physicians also report a meaningful drop in consultations where prior imaging from another institution was unavailable — previously common enough to force repeat investigation at the hospital's own facilities.

Phase 3 outcomes remain qualitative pending completion of the six-month trial.

---

This engagement is published with the client's identity withheld. If you are working under similar constraints — real records, real accreditation, and no permission to send any of it to a cloud API — that is the problem this was built for. Tell us what you are running, or read more about our healthcare practice.