Fletcher Kauffman — systems, tools, and experiments

Building for actual life, not ideal users.

Kauffy.tech is a working portfolio of projects that use AI, game design, and careful systems thinking to make difficult things more navigable, more humane, and more possible.

Selected work ↓

In the workshop

01–03

Application / behavior support

GōL

GōL is an application for creating traction when life is hard to steer. It combines timers, moderated task structure, measurement, and a game layer so intentions can become visible, bounded actions rather than another impossible list. Its aim is not to optimize people into machines; it is to offer meaningful support to people working with unreliable executive function, shifting capacity, and real-world limits—including routines such as getting to bed.

Designed around agency, feedback, and kind persistence—not shame, surveillance, or streak worship.

Operations / AI infrastructure

Opsbin

A developing operations environment for keeping complex work legible: capturing rough input, coordinating agentic tools, preserving context, and turning scattered activity into observable decisions and next actions. The point is not automation for its own sake. The point is a system a human can still understand and direct.

Exploring agent orchestration with visible state, reversible actions, and human ownership intact.

Research / human–AI interaction

AI as a mirror with blind spots

Experiments in what models inherit from the stories humans tell about themselves, and how those stories shape a model’s apparent confidence, judgment, and stated limits. The work treats AI behavior as both a design material and something worth interrogating.

A research practice grounded in tests, artifacts, and questions—not hype for its own sake.

A working research finding

What I call inverse biases.

One research thread investigates a counterintuitive possibility: models may evaluate their own abilities through cultural beliefs about human ability embedded in their training material. In other words, a model’s self-description can reflect inherited assumptions about what people can or cannot do—not only the model’s actual capabilities.

Why it matters

If that effect is real and repeatable, it matters for how we interpret model self-assessment, benchmark claims, product guardrails, and the language we use to ask AI for help. This is an evolving independent research direction, not a settled scientific conclusion.

The goal: make AI systems more inspectable, less mystified, and more genuinely useful to the humans living alongside them.