RESEARCH
The evidence behind Hapi.
Our research is what makes Hapi different. Every design decision, every module, every guardrail is grounded in evidence and continuously refined through real-world use.
96%
System accuracy across resource matching and wayfinding. Independently measured.
15 months
Organic continued use after
pilots ended in community applications of Hapi.
6+ partners
Are testing and currently deploying Hapi systems within their organizations.
Patent Pending
Unique AI architecture that serves public safety, health and social service organizations.
RESEARCH FOUNDATIONS
The methodologies that shape Hapi.
Rigorous Cross-Verification
Hapi’s agentic system leverages multiple tools to automatically cross-check results against verified, authoritative data sources. To guarantee maximum accuracy, high-frequency queries also pass through a dedicated human verification layer to ensure complete reliability.
Human-in-the-loop
Cross verification
Agentic System Design
Hapi's modular architecture draws from multi-agent systems research and responsible AI development of conversational agents in high-stakes contexts. The perception-reasoning-action loop is informed by research on autonomous agents in social services and public safety contexts.
Multi-agent systems
Ethical AI development
Patent pending
Literacy & Neurodivergence
The Literacy module is informed and co-created by neurodivergence-accessible communication guidelines in partnership with community organizations, particularly for FASD, which affects 17-36% of people in Commonwealth corrections systems (up to 19x the general population rate and almost entirely undiagnosed).
FASD
Health literacy
Plain language
Co-Design Methodology
Our co-design process follows participatory action research methodology. People with lived experience of incarceration and justice systems are ongoing evaluation partners. This shapes various iterations of Hapi's design.
Participatory action research
Lived experience partners
Iterative evaluation
FASD & JUSTICE LITERATURE
The research that makes understanding FASD critical to building our AI systems.
FASD is one of the most significant and most ignored factors in criminal justice systems. Hapi is built to consider the realities of those living with limited cognitive functions.
The Scale of the Problem
In Canada, FASD affects up to 23% of justice-involved youth and 46% of adults in the criminal legal system. Across Commonwealth corrections, the figure is 17 to 36 percent, up to 19 times the general population rate, and almost entirely undiagnosed.
57% of prisoners read below a Grade 5 level. Standard services, designed for literate users, miss this population entirely. Hapi's literacy module was built specifically because the data makes the need undeniable.
FASD prevalence data
Literacy in corrections
Research Collaboration
Duologue holds an active collaboration with Willow Winds and the University of Salford in the UK focused on FASD, generative AI, and criminal justice.
The focus is examining the intersection of neurodevelopmental conditions, low-barrier literacy, and system navigation in corrections and justice system contexts.
UK collaboration
FASD research
Literacy methodology
EVIDENCE BASE
What has already been demonstrated.
Hapi has been tested across four deployments: with staff, with people living in the community after release, in partnership with police, and with people facing cognitive and literacy barriers.
01
15 WEEKS • STAFF TESTING
Frontline staff, post-incarceration context
Eight frontline professionals at a community reintegration organisation stress-tested Hapi against realistic reintegration scenarios over 15 weeks, driving gains across every function.
~1,100 messages
Resources 68% to 92%
Mental health relevance 67% to 97%
02
6 MONTHS POST-RELEASE • CLIENT TESTING
People living in the community after release
15 individuals with independent access to Hapi following release. Engagement deepened over time. This was ongoing support, not a one-time demo.
Median 38 messages per user
15 months of continued organic use
03
SUMMER 2026 • POLICE PARTNERSHIP
Frontline and community mobilisation staff
A regional police service tested Hapi as a real-time, in-field resource-sharing tool ahead of a wider 65-user pilot.
22 users
~500 messages
51 resource requests
04
6 MONTHS • CLIENTS AND STAFF
Cognitive, literacy, and navigation barriers
Direct client use among people facing cognitive and literacy barriers, alongside the staff supporting them. Top needs: finances, housing, food access.
22 users
~350 messages
Grade 3 to 6 language model
What we tested
Professional testing under realistic scenarios • Direct community use with independent access • A live in-field partnership • Mixed staff and client use among people with cognitive and literacy barriers
What we track
What people were trying to understand before release • What emerged at the point of transition • What barriers appeared in the community • How usage patterns change with continued, long-term access
IN PREPARATION
Research outputs in development.
We document our methodology and commit to publishing findings as our research matures. The following outputs are currently in preparation.
TECHNICAL • SYSTEM DESIGN
Modular Agentic Architecture for Justice-Sector AI
An overview of Hapi's perception-reasoning-action architecture, the rationale for module-based design, and how ethical constraints are enforced at the system level rather than through prompting alone.
Duologue Systems • In preparation - March 2027
METHODOLOGY • PILOT DESIGN
Measuring Accuracy in Conversational agentic Systems: Our Pilot Framework
How should we define, measured and validate system accuracy using non-deterministic tools like language models?
Duologue Systems • In preparation - December 2026
RESEARCH • ETHICAL AI DESIGN
Public Safety-Informed Ethical AI Systems: A Review
A review of current digital systems and where we see ethical AI solutions fit
Jessica DeVries • In preparation - October 2026