How We Built It
Home How We Built MIA Research Team About UNCAPT Partners Request Pilot

How We Built MIA.

Two decades of clinical research. Specialist expertise from the Brain and Mind Centre. Iterative testing with real clinicians. This is how we engineered AI reasoning and knowledge systems you can trust with care decisions.

Poulsen et al., CHI 2025 · Iorfino et al. (in preparation)
Scroll

"We didn't train a chatbot. We encoded decades of specialist reasoning into an AI system that thinks the way clinicians think."

— The MIA Design Philosophy

The Approach

Three inputs.
One intelligence.

MIA was built at the intersection of three distinct knowledge sources — each one essential, none sufficient alone.

Expert reasoning provides the how. The evidence bank provides the what. User testing provides the proof.

The result: an AI agent that matches expert consensus 95% within top-2 choices.

MIA Agent Expert Reasoning Evidence Bank User Testing
Three Pillars

What makes MIA different.

MIA wasn't trained on the internet — it was built on specialist knowledge, curated evidence, and continuous human feedback. Three clinical pillars, each deliberately engineered.

01

Case-Based Expert Reasoning

Specialists didn't just review outputs — they authored the reasoning. Clinicians codified how they identify critical features, structure assessments, and build domain-specific care plans. MIA reasons using a multidimensional clinical framework covering what to assess, how to score it, and when to escalate — mirroring the decision process of an experienced practitioner, not a generic chatbot.

Critical Features Care Plans Assessment Logic Escalation Criteria Biopsychosocial
20+
Years of clinical research
powering MIA's knowledge
BRAIN AND MIND CENTRE
95%
Agreement with expert
consensus (top-2)
VALIDATED ACCURACY
02

Evidence-Based Knowledge Bank

Curated from two decades of Brain and Mind Centre research — clinical guidelines, validated instruments like the IAR-DST, and treatment protocols across care levels 1–5. This isn't a static reference library; it's a living, versioned knowledge base that MIA actively reasons over to ground every assessment, recommendation, and care plan in peer-reviewed evidence.

IAR-DST 8 Clinical Domains 5 Care Levels Treatment Protocols Clinical Guidelines
03

Continuous Expert Feedback

Mental health professionals and individuals with lived experience iteratively shaped MIA — reviewing clinical outputs, correcting reasoning paths, and surfacing edge cases that became safety constraints. This isn't a one-off training run; specialists continuously refine how MIA handles nuanced clinical scenarios, ensuring it improves with each feedback cycle rather than drifting.

Clinician Review Lived Experience Safety Guardrails Edge Cases Iterative Refinement
Continuous expert
feedback cycles
ITERATIVE REFINEMENT
Milestones

From research to product.

The key moments in MIA's journey — from foundational research to a validated clinical AI agent.

2003 – 2023

BMC Youth Mental Health and Technology Program

The Brain and Mind Centre's Youth Mental Health and Technology team conducts two decades of transdiagnostic mental health research, developing clinical staging models and the multidimensional assessment framework that would become MIA's foundation.

12th June, 2023

UNCAPT × USYD Partnership

UNCAPT and the University of Sydney commence proof of concept work — operationalising BMC's clinical expertise as an AI agent on UNCAPT's agentic platform.

2023-2024

Knowledge Encoding & Build

Clinical specialists begin encoding reasoning and evidence into MIA. The Knowledge Bank is curated, expert evaluators authored, and the first version of MIA goes live on the UNCAPT platform.

2024 – 2025

Iterative Testing

Clinicians and individuals with lived experience test MIA across hundreds of scenarios. Continuous feedback cycles refine clinical accuracy, conversational flow, and safety guardrails.

NOVEMBER 2025

Validation

MIA achieves 95% top-2 agreement with expert consensus, validating the system's clinical reasoning across the BMC multidimensional framework for youth mental health.

The Platform

Built on the UNCAPT Agentic AI Platform.

MIA's clinical intelligence doesn't run in isolation. It's operationalised on UNCAPT's purpose-built agentic AI platform — a secure, scalable infrastructure designed to host domain-specialist AI agents across regulated industries.

The platform provides the orchestration layer: the reasoning engine, memory, evaluation harness, and deployment infrastructure. The Brain and Mind Centre provides the clinical knowledge. Together, they form MIA.

Agentic Engine (OODA Controller)

Observe–Orient–Decide–Act loop with tool routing, memory management, evaluation harness, escalation logic, and chain-of-reasoning logging.

Expert Training Platform

Enables subject-matter experts to converse with the agent, edit and rewind thought processes, and provide feedback to produce fine-tuning datasets.

Knowledge Bank Infrastructure

Ingestion, vectorisation, clustering pipelines, contradiction analysis, query engine, and a visual portal to explore the curated knowledge bank.

Secure Cloud Platform

Azure Australian-hosted deployment with co-pilot and autopilot modes, web interfaces, and enterprise-grade security and observability services.

Under The Hood

What specialists built into MIA.

Clinical experts directly shaped MIA's reasoning engine — here's what they contributed.

Specialist-Guided Reasoning

Clinicians provided structured input to ensure MIA's assessments mirror expert judgement.

  • Critical features — Key indicators that must not be missed during assessment
  • Specific care plans — Evidence-based treatment pathways for clinical profiles
  • Assessment structure — Logical flow and priority of clinical questioning
  • Multidimensional framework — Biological, psychological, and social domains

Curated Evidence & Resources

Two decades of research from the Brain and Mind Centre powers MIA's knowledge bank.

  • Clinical guidelines — National and international best-practice frameworks
  • Research publications — Peer-reviewed transdiagnostic assessment studies
  • Assessment instruments — Validated tools including IAR-DST (8 domains)
  • Treatment protocols — Matched interventions for care levels 1–5
Research Foundation

Built on decades of clinical evidence.

MIA integrates the research output of the Brain and Mind Centre's Youth Mental Health and Technology team — spanning transdiagnostic models, staging frameworks, and measurement-based care.

20+

Years of research

8

IAR-DST domains

5

Care levels mapped
Poulsen et al., CHI 2025 User-centred design of AI clinical agents
Iorfino et al. (in preparation) Validation of AI-driven IAR-DST scoring
MJA — Right Care, First Time Personalised, measurement-based youth mental health care
BMC Transdiagnostic Framework Staging model for clinical need levels 1–5
Iterative Process

Refined through human feedback.

MIA wasn't built in a single pass. Each capability went through repeated cycles where clinical specialists interacted directly with the agent — observing its reasoning in real time, identifying where logic broke down, and providing structured corrections that fed back into the system.

This isn't passive annotation. Through the training platform, experts converse with MIA, edit and rewind its thought processes, and give positive or negative feedback — producing the datasets that progressively sharpen clinical accuracy.

This cycle runs continuously — every edge case surfaced becomes a permanent design constraint, not a one-off fix.

1

Design Scenarios

Specialists craft clinical scenarios targeting edge cases — ambiguous presentations, comorbidities, high-risk indicators — across all 8 domains.

2

Observe & Converse

Experts interact with MIA in real time, tracing its reasoning chain to see exactly where clinical logic holds or breaks down.

3

Edit & Correct

Specialists rewind decisions, edit thought processes, and provide structured feedback — each correction producing fine-tuning data.

4

Validate & Repeat

Updated model benchmarked against multi-expert consensus. If thresholds aren't met, the cycle restarts with new scenarios.

The Result: MIA.

A clinical intelligence agent that matches expert consensus 69% exactly and 95% within top-2 — ready to transform mental health assessment, triage, and care planning at scale.