Hard to steer
Behavior depends on model training, prompts, and hidden orchestration.
Research
Psychological AI and narrative systems
Trifrons researches how artificial intelligence can interpret, structure, and support autobiographical meaning-making. These prototypes explore narrative analysis, psychological AI, and interpretable systems.
AI-powered analysis of autobiographical narratives
Subconscious Structure Mapping
AI tools are increasingly used for coaching and self-reflection, but many remain:
Opaque to inspect
Interpretations cannot be easily traced or checked.
Weakly grounded
Assumptions, data sources, and intended use are often unclear.
SSM aims to address this through a modular, interpretable architecture with 3 major layers:
- SSM System — explicitly separates Control (user input analysis, LLM orchestration and reply generation) and Memory (modeling session and user state and applying analysis diffs deterministically).
- Psy KB — a psychological knowledge base consisting of an Analytic Ontology (a causal model of what to extract from the text, a theoretical hypothesis) and a Narrative Graph (a series of acts and transitions to elicit the user to produce output from which the concepts may be extracted).
- UX/UI — a user experience and interface layer with pluggable AI personas and narrative scenes to realize the narrative graph and provide engaging journeys for the user.
SSMv0 Prototype
An early burnout-prevention AI system that informed the SSM architecture and design. The original psychological knowledge base, act-based structure and tone of voice were tested and improvements identified.
Open SSMv0 live prototype External link — opens in a new tabQodex Narrative Atlas
An exploration of a psychological narrative atlas built from grounded symbolic material such as astrological aspects and tarot card pairs. The approach proved incomplete because symbolic material alone did not capture actor and setting context.
Open Qodex fragment External link — opens in a new tabAstro-tarot Scene Generation
A prototype for generating narrative scenes from astrological transits and tarot cards as a possible Qybala v2 experience. People rating horoscope-style situation–behavior hypotheses could potentially inform a tentative psychological profile.
Explore sample scenes External link — opens in a new tabPsychodynamic narrative model prototype · APRIL-JUNE 2024
AMKOB Narrative Analysis
AMKOB adapted JAKOB External link — opens in a new tab , an established psychodynamic narrative method for analyzing conflict developed by Prof. Dr. Brigitte Boothe External link — opens in a new tab who advised this project. AMKOB translated this method into software and extended it with structured frame annotations, self-agency statistics, and a coreference-based Object Graph.
The prototype transformed short stories into inspectable structures, but development was paused pending a fundamental upgrade of the psycho-linguistic resources and because of the limitations of the offline LLMs available at the time.
Narrative actions and agency
- Verb coding — verbs are treated as central indicators of action and conflict, using the published JAKOB classification.
- Frames and JAKOB annotations — structure how actions, actors, and conflict patterns are represented.
- Self-Agency Stats — build on these annotations to examine patterns of agency across the narrative (Akteurenschicksal).
Object Graph
- Objects become nodes — including characters and relevant props.
- Relationships are extracted — through coreference analysis using a custom model and heuristics over NLP-processed text.
- Graph structure can then be analyzed — to explore possible psychosocial patterns around the protagonists.
EU TRUSTING PROJECT · PUK ZÜRICH · JANUARY–MARCH 2024
Text coherence in psychotic speech
During a three-month research appointment at the Psychiatric University Clinic Zürich, in the Homan Lab, I worked on the EU TRUSTING project, reviewing and experimentally evaluating computational methods for detecting clinically relevant changes in psychotic speech.
Methods reviewed
- Discourse coherence models — including DiscoScore and rhetorical structure approaches.
- LLM-based evaluation — using language models to assess coherence and discourse organization.
- Clinical applicability — comparing how these methods could be applied to existing psychiatric interview data.
Output
The work resulted in a technical literature review covering the state of coherence modeling, major methodological limitations, and promising directions for future evaluation.
Collaborate on research
I am open to academic, clinical, and technical collaborators interested in psychological AI, narrative systems, and interpretable human-centered research.
Explore a research collaboration External link — opens in a new tab