Demo / engineering overview: Metrics on this page reflect internal test harness runs and may change. This is not clinical validation or medical advice. Please avoid entering identifying patient information.

Technical Architecture

Technical Details (Demo)

This page summarizes Vitruviana’s architecture, model-routing approach, and internal evaluation notes for specialty-tuned clinical workflows. It is an engineering overview and should not be interpreted as clinical validation.

Demo build

Vitruviana Engineering

Internal test harness

System snapshotcorr_id: demo-23A9

Input

Voice + text intake

Safety gate

Clinician review required

Escalation + routing

Model router

Gemini 3 Pro + GPT-5.2

ReasoningStructured

Artifacts

SOAP, orders, AVS, MDM

Schema-validated outputs

Trace preview

00:00.042task.routestructured extraction

00:00.218model.selectGPT-5.2

00:01.142artifact.generatepacket + codes

System snapshot

Demo

Engineering Overview

This page summarizes architecture + internal evaluation notes (not clinical validation).

Structured

Outputs

Schemas + validation for consistent handoffs and clinician review.

Guardrails

Safety posture

Explicit scope limits, risk flags, and clinician-in-the-loop review.

Router

Model selection

Routes tasks to the best model for reasoning vs structured generation.

Internal

Test harness

Scenario-based regression checks; metrics vary by configuration and run.

Multi

Specialties

Explored on representative demo scenarios across specialties.

System layers + contracts

Experience layer

Patient intake, clinician cockpit, and specialty demos.

Patient voice intakeClinician cockpitRadiology workbenchPsychiatry + screening modules

Workflow services

Session orchestration, routing, and structured packet generation.

Session managerSafety routingSchema validation

Model router

Routes tasks to reasoning vs structured generation models.

Gemini 3 ProGPT-5.2Fallback rules

Clinical services

Insight engine, note generation, orders, and billing hints.

InsightsNote generatorOrders + AVS

Telemetry + QA

Acceptance rates, latency, and audit trail capture.

Draft deltasSafety triggersLatency metrics

Request-to-draft flow

Capture

Voice/text intake or clinician context.

Normalize

Schema validation and safety checks.

Route

Model selection for task type.

Generate

Draft outputs + citations.

Review

Clinician edits + acceptance logging.

Model routing

Gemini 3 Pro

Primary - Clinical Reasoning

Avg latency: 28.3s • Task share 58%

GPT-5.2

Fast - Structured Tasks

Avg latency: 3.0s • Task share 42%

Routing studio

Structured extraction

Convert transcripts into schema-validated packets.

Route: GPT-5.2

Clinical reasoning

Synthesize assessment and treatment plan drafts.

Route: Gemini 3 Pro

Patient communication

Simplify clinician notes into patient language.

Route: GPT-5.2

Safety routing

Detect escalation signals and route for review.

Route: Safety router + model

Engineering validation

InsightEngine

Evidence-based clinical insights with guideline citations

100% success • 31s avg • 8.6 insights/patient

MedicationReconciliation

Drug safety analysis with polypharmacy support

100% success • 33s avg • 12.3 interactions detected

NoteGenerator

Complete SOAP notes with 4/4 section compliance

100% success • 27s avg • 509 words avg

PreVisitInterviewer

Natural patient communication with clinical data capture

96% success • 834ms avg • 89.2% extraction

Specialty coverage

Cardiology

16 scenarios

Pulmonology

4 scenarios

Emergency Medicine

2 scenarios

Primary Care

4 scenarios

Endocrinology

4 scenarios

Nephrology

2 scenarios

Psychiatry

2 scenarios

Oncology

2 scenarios

Public research references

Long-context

Long-context models enable richer clinical context

Google DeepMind / Google Research (public materials)

Multimodal

Med-Gemini: Advancing Multimodal Medical Capabilities

Google Research

Large context

1M Token Context Window for Full Patient History

Google DeepMind

Healthcare models

MedGemma: Healthcare AI Foundation Models

Google for Developers

Want to see the workflows in action? Start with the live demos, then come back here for architecture details.

Engineering blueprint

Services, contracts, and traceability

Each service exposes a narrow contract so we can validate outputs, capture edits, and explain every draft.

Session Orchestrator

Contract

session.start → packet.normalize

Manages intake state, assigns correlation IDs, and tracks draft lifecycle.

Safety Router

Contract

packet.normalize → safety.evaluate

Detects red flags, enforces stop-the-flow, and escalates crisis pathways.

Model Router

Contract

task.route → model.select

Routes reasoning vs structured tasks across GPT-5.2 + Gemini 3 Pro.

Artifact Generator

Contract

model.response → artifact.generate

Builds SOAP, orders, AVS, and clinician summaries with schema validation.

Telemetry + QA

Contract

artifact.review → metrics.log

Logs latency, acceptance rate, edit deltas, and safety triggers.

Trace snapshot

corr_id: demo-23A9

00:00.012session.startPatient voice intake (demo)

00:00.245packet.normalizeHPI + ROS + meds validated

00:00.418safety.evaluateNo emergent flags

00:00.572task.routeStructured output → GPT-5.2

00:01.944artifact.generateSOAP + orders shortlist

00:02.108metrics.loglatency=2.1s, draft=ready

Output contracts validated • edits captured • clinician review required

Architecture map

System layers + contracts

A modular stack designed for clinician review, safety routing, and measurable outcomes.

🎙️

Experience layer

Patient intake, clinician cockpit, and specialty demos.

▾

Patient voice intake

Clinician cockpit

Radiology workbench

Psychiatry + screening modules

🧭

Workflow services

Session orchestration, routing, and structured packet generation.

▾

Session manager

Safety routing

Schema validation

⚡

Model router

Routes tasks to reasoning vs structured generation models.

▾

Gemini 3 Pro

GPT-5.2

Fallback rules

🧬

Clinical services

Insight engine, note generation, orders, and billing hints.

▾

Insights

Note generator

Orders + AVS

📊

Telemetry + QA

Acceptance rates, latency, and audit trail capture.

▾

Draft deltas

Safety triggers

Latency metrics

Runtime pipeline

Request-to-draft flow

Every step is logged to support clinical review, QA, and post-hoc evaluation.

🎙️

Step 01

Capture

Voice/text intake or clinician context.

🧾

Step 02

Normalize

Schema validation and safety checks.

🧭

Step 03

Route

Model selection for task type.

⚡

Step 04

Generate

Draft outputs + citations.

✅

Step 05

Review

Clinician edits + acceptance logging.

🏆Public research references

Model + architecture context

Selected public references related to long-context and multimodal foundation models. This is not an endorsement or clinical validation.

Long-context

Long-context models enable richer clinical context

Modern foundation models can process substantially larger contexts, enabling review of longer histories and more complete documentation drafts (deployment-dependent).

Google DeepMind / Google Research (public materials)•2024–2025

Multimodal

Med-Gemini: Advancing Multimodal Medical Capabilities

Multimodal medical foundation models continue to improve on benchmark tasks across imaging + text domains.

Google Research•2024–2025

Large context

1M Token Context Window for Full Patient History

Large context windows can support longer chart review and multi-source synthesis (depending on model and deployment constraints).

Google DeepMind•2024–2025

Healthcare models

MedGemma: Healthcare AI Foundation Models

Domain-focused model families aim to improve medical reasoning and understanding across modalities.

Google for Developers•2024–2025

Engineering validation notes

High-level highlights from internal demos and test harness runs (not clinical validation).

🛡️

Demo

Engineering Overview

This page summarizes architecture + internal evaluation notes (not clinical validation).

📋

Structured

Outputs

Schemas + validation for consistent handoffs and clinician review.

💊

Guardrails

Safety posture

Explicit scope limits, risk flags, and clinician-in-the-loop review.

🎯

Router

Model selection

Routes tasks to the best model for reasoning vs structured generation.

🔬

Internal

Test harness

Scenario-based regression checks; metrics vary by configuration and run.

🏥

Multi

Specialties

Explored on representative demo scenarios across specialties.

Hybrid architecture overview

Routing separates deep reasoning from structured generation; specific latencies vary by configuration.

🧠

Gemini 3 Pro

Primary - Clinical Reasoning

28.3s

Avg Latency

58%

Task Share

Routed Tasks:

DiagnosisMed ReconciliationRisk AssessmentEvidence Synthesis

⚡

GPT-5.2

Fast - Structured Tasks

3.0s

Avg Latency

42%

Task Share

Routed Tasks:

SOAP FormattingJSON ExtractionPatient CommsQuick Lookup

Routing studio

Router + artifact mapping

Select a task type to preview the routing path and the artifacts produced for clinician review.

Task types

Route previewInteractive

Task

Structured extraction

Convert transcripts into schema-validated packets.

→

Router

GPT-5.2

Best-fit model for task type

→

Artifacts

Structured packetICD-10 candidatesMeds + allergies

Every artifact is schema-validated and reviewed by clinicians before use.

Trace logsession: demo

Real clinical scenarios tested across 8 medical specialties

❤️

Cardiology

16 scenarios tested

🫁

Pulmonology

4 scenarios tested

🚨

Emergency Medicine

2 scenarios tested

🏥

Primary Care

4 scenarios tested

🔬

Endocrinology

4 scenarios tested

🫘

Nephrology

2 scenarios tested

🧠

Psychiatry

2 scenarios tested

🎗️

Oncology

2 scenarios tested

Statistical Validation

Model Comparison

Latency Difference:25,225ms (p=0.001)

Effect Size (Cohen's d):3.71 (very large)

95% Confidence Interval:[21,009ms, 29,441ms]

Production Metrics

System Reliability:95.0%

Error Rate:2.0%

Total API Cost:~$0.024

Explore the Demos

Try the live workflows and see how clinician review, structured outputs, and model routing fit together.

Launch Clinical Cockpit Try Patient Demo

Explore the Demos Open Clinician Cockpit

Engineering overview • not clinical validation