Amigon

My role

Concept, design and implementation, solo

The problem

Structured mental-health questionnaires are accurate but tedious, so people stop filling them in. Free-form journalling is easy to keep up but produces nothing a clinician can work with. Amigon sits between the two: a short conversational check-in that still produces structured, standards-compliant output.

How it works

• Four validated screening instruments covering depression, anxiety, wellbeing and affect, rather than inventing new items.
• A voice check-in. Instead of only answering scales, people can speak. Affect is estimated from two channels, acoustic features of the voice and a language model reading what was actually said, and the two are fused with confidence weighting rather than one overriding the other.
• A safety layer that screens responses for indications of acute risk on a staged model. When it triggers, the app stops what it was doing and shows one screen with two buttons: the emergency number and the crisis helpline. It is a signposting mechanism, not a diagnosis, and the app says so.
• Every measurement becomes an HL7 FHIR R4 Observation, coded against ICD-10-GM, SNOMED CT and LOINC.

Stack

Flutter and Dart, Firebase with Firestore, Auth and Cloud Functions, Python for the acoustic analysis pipeline, an LLM for the linguistic channel, HL7 FHIR R4 with ICD-10-GM, SNOMED CT and LOINC.

What I take from it

Building the FHIR layer took a large share of the implementation time and gave the app itself precisely nothing. No feature depends on it and no user sees it. I did it anyway, because health data that only exists inside one application is worthless the moment that application goes away, and applications always go away. A score sitting in a proprietary database is a number. The same score as a LOINC-coded Observation is something another system can still read in ten years.

Outcome

The SUS score of 86.8 is the headline, against a published average of 68. The free-text answers taught me more. Several people described the voice check-in as easier than the questionnaires precisely because it did not feel like being assessed, which was the entire hypothesis and also the part I was least sure about.