A few things we're still figuring out

fella already does one specific job well: point it at a folder nobody organized for it, and it answers real questions about that folder with a real number behind every answer. Every bet below is really the same bet, made four different ways: get dramatically better at that one job, not add new ones to it.

The closest thing that already exists is the kind of natural-language analytics a company pays a data team to build on top of a warehouse: ask a plain question, get a trustworthy answer with the work shown. fella's version doesn't get a warehouse, a data team, or months of setup, it has to work out the shape of your own messy folder on the fly and still hold to the same bar of proof. That's a harder version of the problem, not an easier one, and it's the actual target: something closer to enterprise-grade analysis than a chat app, running entirely on your own machine, pointed at whatever your folder actually looks like.

That's a different kind of “more” than most of the industry is chasing. It's not more tools, more autonomy, or more places fella can reach, every bet below stays inside the same one folder, read-only, verified shape that already exists. It's the same bounded job, done at a depth most personal tools never attempt.

Right now, fella needs a model to actually answer your questions: either one you install yourself (free, but you have to set it up), or one you sign into somewhere else. The bigger goal is fella coming with its own small, free model built right in, so the very first time you open it, it already works. No account anywhere, no key to paste in, ever.

That model won't try to know everything. It'll be built to do one or two specific jobs really well, figuring out which of your files actually matter for a given question, say, rather than trying to be a do-everything brain. Small and genuinely good at one thing, not big and vaguely okay at everything.

The technical version: picking the right files, cleaning up messy data, and writing its own queries.

fella already double-checks its own answers before showing them to you. It re-runs the actual calculation and flags anything that doesn't add up, rather than just trusting itself. We want to make that habit even harder to break, and eventually let other people building their own AI tools use the same checking system, instead of everyone building a worse version of it on their own.

The technical version: a standalone checking library, and holding our own rules to the same standard.

Everything works great when we test it on a tidy folder. Real folders aren't tidy: years of files, all different sizes, nothing named the same way twice. We deliberately tested fella against an actually huge, messy folder, and found two real problems: it gets slower than it should, and it can get confused about which file you mean if you don't spell it out. We're fixing both for real, not just papering over them.

The technical version: why it slows down at that size.

We can't personally find every strange way a real person's files break things, or test every AI model ourselves. We'd rather open that up: let anyone check how well a given model actually performs with fella, and let anyone who finds a genuinely weird real-world mess add it to the pile of things fella gets tested against.

The technical version: a public leaderboard for models, and an open list of real messy-data traps.

fella works today because it's small, fast, and doesn't ask you to reorganize your files to fit it. Every bet above only counts if it survives contact with that rule. A built-in model that's actually fast and tiny, not a bigger dependency wearing a feature's name. Checking that stays instant, not a system you sit and wait on. Handling a bigger folder without slowing down, not just tolerating it. Opening things up for contributors without opening up complexity for the person who just wants to ask a question.

If any of these ideas end up making fella heavier, slower, or harder to actually use, we'd rather cut the idea than break the thing that got you to install it in the first place.

Short, dated notes as any of this actually moves, not a running commentary, just what actually happened.

• 2026-09-13 This page went up, alongside the real issues linked above. Nothing built yet. This is the starting line, not a progress report.