Michael Kupermann
Advisory work and development

Methods, tools
and open projects.

My work includes jointly developed methods, tools for data and AI, and open research and teaching projects. FireScore brings employees' perspectives into the planning of change. The other projects make selected methods and technical approaches available for examination. Each description sets out its purpose and limits.

Projects

Purpose, approach and limits.

Framework

FireScore

An AI programme changes the work people do. Understanding which tasks they value and where they would welcome relief belongs in its planning. I developed FireScore with Silke Rengstorf to bring that perspective into the discussion. Its four dimensions examine the value of doing a task oneself, professional identity, resistance to automation and whether the work energises or drains.

An anonymous survey provides a basis for conversations with management, employees and their representatives. The findings help frame discussions about future responsibilities, concerns and support. They reflect the respondents' own assessments and do not predict the success of a technology programme.

FireScore and its approach ↗
Method

Kupermann Decision Partner

A structured working method for using a language model in complex decisions. It covers the problem, criteria, options, focused investigation, challenge and subsequent review. The person retains responsibility for objectives, trade-offs and the decision.

The repository contains a skill, templates, primary research references and a fictional worked example. It draws on established problem-solving methods. The method's additional effectiveness has not been demonstrated by a comparative study.

Read the article · Repository ↗
Local AI memory

Throughline

A vendor-neutral memory store for use with multiple AI assistants. PostgreSQL and pgvector hold the data on infrastructure controlled by the user. The design addresses continuity between conversations and the ability to retain information independently of one assistant provider.

This is beta software. Integration support and deployment requirements are documented in the repository. Keeping the memory store local does not determine where a connected language model processes its prompts. That depends on the selected model and configuration.

Repository ↗
Structured-data enquiries

Souprise

A local question-answering system for invoices, orders and customer records. Its design separates calculation and retrieval from language generation. Code handles totals and averages, while the model supports the presentation of the answer.

The project explores a bounded approach to answering from known data and declining unsupported requests. Published tests cover specific questions and synthetic data. They do not establish correctness for every query, dataset or configuration. Deployment still requires review of access controls, data coverage and failure handling.

Repository ↗
Exploratory research

vibrasim

A simulation project investigating whether a simple abstract substrate can exhibit selected memory-like properties. Its terms, including vibration, electron and atom, name simulated nodes and relationships. They are not claims about physical particles or a model of their physics.

A recorded run on 31 May 2026 covered twenty pre-registered experiments with controls. The published result was negative against the stated acceptance criteria. That result describes those experiments, not every later version of the project. The associated research workflow is also available as a separate repository.

Repository ↗ · Research workflow ↗
Teaching project

A German-language transformer from scratch

An educational implementation intended to make a transformer's components readable. Six tutorials cover the model architecture, attention, feed-forward layers, German tokenisation, training and data collection. A small dataset builder and training script accompany the explanations.

The purpose is understanding the implementation. It does not establish that a model trained with this project matches production language models in capability or reliability.

Repository ↗
ML analysis tool

EquiML

A command-line and Python tool combining fairlearn, SHAP and scikit-learn in a workflow for loading data, examining bias metrics, training and comparing models. Outputs include an HTML report and metrics in JSON.

The work makes the trade-offs between selected fairness measures and predictive performance explicit. The choice of measures and the interpretation of the results remain part of the analysis. A report is not a general certification that a model is fair.

Repository ↗
From method to a decision

The application determines the work.

A project's relevance depends on the data, people and operating constraints around it. For a business decision, the useful starting point is the problem to be resolved and the evidence required to proceed.

Discuss the application → · Read the Insights articles →