flowfab

Independent applied AI research

One human, a rotating cast of AI collaborators, no meetings.

FlowFab is where I study the questions that come up while building my own systems: a game master with long memory, agent architectures, and video made by language models. Each note says how it was tested and where it stops being true.

About the lab
Status flowfab.io
~> flowfab status
staff1 (me)
meetings0 this week
hoursevenings and weekends
agents0 on payroll (they work for tokens)
side projectstoo many
~>

01 Research lines

L1

Memory and coherence in narrative agents

How does an AI game master remember a whole campaign without lossy summaries, and keep track of what each character knows?

L2

Agent architecture

How do you design agents with continuity and autonomy, and compare design proposals from several models fairly?

L3

Media production with LLMs

How far can an AI pipeline go in producing good explainer animation with little human intervention?

02 The bench

Personal projects. They are not research, but most of the research questions start here.

B1

Silverlight

An AI game master for long solo campaigns. I wanted a GM who never cancels the session.

B2

Any Given Thing

Animated explainer videos on YouTube, made with an LLM pipeline. Because I have a day job.

@anygiventhing ->
B3

PrintStream

A manager for small 3D printing farms, starting with mine. In development.

B4

3D printing

One printer, a lot of PLA, and the occasional first layer that refuses to stick.

03 Method

FIG. 1

Every note follows the same seven steps.

  1. 01Question
  2. 02Context
  3. 03What I tried
  4. 04How I measured
  5. 05Result
  6. 06Limits
  7. 07Next step