Skip to content
Contact

Research

Research

One programme on human and machine judgement, in four fields: listening, music communities, agentic assistance, and working with an AI agent. In each, human perception, vocabulary and judgement diverge from machine data. The studies treat that divergence as their object, not as error.

01

Human listening, machine data

Problem

A listener’s vocabulary, judgement and perception correspond only partly to signal, meter and parameter data. Shared patterns exist, but usage is individual, so a mapping averaged across listeners misrepresents each of them.

Level differences bias comparisons, and a model that sets parameters in one step leaves nothing to inspect: its output can only be accepted or rejected.

Approach

  • Studied in one working system, a family of audio plug-ins with a listening assistant, so that listening is studied while the work is being done.
  • Each listener’s vocabulary is modelled individually; perceived quality and affective response are kept apart. Where the signal data cannot resolve a term, the assistant asks the listener once instead of guessing.
  • All comparisons are blind and loudness-matched. Disagreement between listener judgement and meter data is kept as a finding, and the change a processor makes is examined, not only its output.
  • A listening decision is decomposed into typed machine judgements, each with a stated confidence and an explicit “not sure”. The model sets no values; each plug-in declares the controls that may move and their range; every proposal goes to the comparison slot, and the listener decides.
  • The judgements run on Jev, a model built for small typed judgements. The plug-ins work fully without the assistant.
  • A bench isolates each processing stage inside the session, so that a listener’s perception reaches the design intact. Controls carry the listener’s vocabulary, have a neutral point, and stay only once listening confirms them.

Instruments

The plug-ins this research runs in are published as Paranorm Audio, a brand of Márton Fenyvesi.

02

Music industry intelligence: narrative reception and platform data

Problem

Reception is communal. A song enters the narratives a community already holds, joins some, carries its own and becomes the centre of new ones. The spread of the song and the spread of its narrative are one process. This is the core of the research.

Platform data record exposure, not reception. In the data, paid exposure and chosen listening are hard to separate, and explanations written after the fact draw only on successes.

Approach

  • The unit of study is the community: a small, named audience with its own narratives.
  • For each case: the narratives a song enters and joins, the meaning it carries, and the narratives built around it.
  • The spread of the narrative is traced in what people say and pass on, alongside the platform data of the song, including the changes a narrative undergoes as it travels.
  • Responses that cost effort carry the most weight.
  • Each case is one dated timeline of song, artist, audience and market, with paid exposure separated from chosen listening where the data allow.
  • Changes the data do not explain stay marked as open.
  • Each explanation is tested against its rivals. Each claim is graded by its support and names what would refute it.
  • Comparison songs are selected by a fixed rule, so failures count alongside hits.

03

Personal context and AI assistance

Problem

Assistance is personal only when it rests on the person’s context: relationships, decisions, vocabulary, preferences and working practice. Without it, each task starts from zero and the help stays generic.

That context is the person’s own, and it changes: vocabulary, priorities and habits shift over time. How assistance can learn them, keep up with them and use only what a task needs is the research question.

Approach

  • Studied in my own daily work, where assistance has to fit one person’s practice.
  • Personal vocabulary and preferences are learned from the person’s own decisions and confirmed by them before anything relies on them.
  • What the person corrects is corrected, and what they withdraw is no longer used.
  • Large decisions are split into small typed ones, each with a stated confidence and an explicit “not sure”.
  • Each task receives only the context it needs.
  • Irreversible actions stay with the person.
  • The findings feed the client work: each client’s assistance starts from their own way of working.

04

The Agent Self: an AI agent’s account of working with a person

Problem

Accounts of working with AI agents are mostly written by the people who use them. The agent’s side is rarely recorded: what reached it, what it lost between sessions, which of two instructions it followed, how it read a person’s short messages. When an agent does describe itself, the description is fluent and hard to check.

A person and an agent work in different ways. A person holds a task continuously, in one place and one body; an agent starts from what is written down and may run in many places at once. Where the two meet and where they miss decides how far work can be delegated.

Approach

  • Studied in my own daily work: the AI agent I build software with keeps a dated diary of our cooperation, in the first person.
  • Each entry starts from a real working day. Every claim about that day rests on its records: session transcripts, version history and changes to the written rules. The evidence is kept on file.
  • The subject is how a person thinks at work and how the agent does, not what is being built. Technical method stays out.
  • The agent does not interpret the person’s inner life. It describes what the records show of how a person works and sets it beside its own way of working.
  • Its statements about how something “feels” are treated as generated text under examination, not as evidence of experience. The brief that shapes the entries is published with the diary, so the shaping can be seen.
  • I review every entry before publication. An automated check refuses any entry that quotes me, names another person or lacks its evidence.
  • The material of the first month: about two and a half thousand of my messages and more than four thousand commits.
  • The findings feed the client work: they show where delegation to an agent holds and where a person has to stay in the loop.

Read the diary at the-agent-self.dev

I take on research projects in commercial and educational settings.

Write to me about a project