We are a research-led company. We build agentic coding tools, finetune our own models, and study what happens when the person writing the prompt is dyslexic — starting in the classroom.
We believe the tools that write software should work for every kind of mind. That belief decides what we build and what we study.
We train our own models for software tasks, so the tools we ship are tuned to the problems developers actually hit rather than to leaderboard scores.
We work on making coding agents do more with less: tighter loops, less wasted context, and output you do not have to reread three times to trust.
We prototype in the open and fold what we learn back into the products. Research that never reaches a user is not much use to anyone.
Building AI that works for how people actually write
The hard part is not reading the answer. It is writing the question. Getting a good result out of a language model increasingly means typing a long, carefully spelled prompt. Dyslexic people process written language differently, so they write differently, and these models were not trained to expect that. The same holds across neurodiversity wherever writing is where the difference shows. Most accessibility work in AI has looked at the output. We think the real barrier sits at the input, and that it is a blocker worth removing rather than a cost worth reporting.
What we are working on:
If you work in this area, or you are dyslexic and want to help us test, get in touch.
Tokenizers are built around conventionally spelled text. Spelled the usual way, a common word often costs a single token. Spelled the way it sounds, it shatters into three or four. research is one token. A common misspelling of it is three.
So a dyslexic user pays more for the same request, fills the context window faster, and reaches usage limits sooner, all for expressing the same idea. It is a per-keystroke tax that nobody has priced.
Dyslexic users are the clearest case, but the same tax falls on neurodiverse people generally wherever writing is where the difference shows. Measuring it is not the point. The cost is evidence of a blocker, and removing those blockers is what we are for. Nobody should have to spell a question correctly to get a good answer to it.
For many dyslexic learners, computing is the subject that works. We would like it to stay that way.
For a lot of dyslexic people, computing is the part that already works. Code is structured, it is read in small pieces, it answers back immediately, and it asks for far less prose than almost any other subject. Plenty of dyslexic programmers will tell you that programming came easier to them than reading and writing ever did.
Large language models change that bargain in both directions. They genuinely help when writing is hard: they draft, they summarise, they explain something a second time without sighing. But they also move the entrance. Getting a computer to do something increasingly means writing a careful paragraph of prose first, which puts the subject that was accessible back behind the skill that is not.
That is why the classroom work and the research are one project rather than two. The training data behind these models was not built on the way neurodiverse people write, so a model has only ever learned one narrow idea of what a well-formed question looks like. That risk sits inside the model, not inside the child, and removing it is the work.
Computing is where a lot of dyslexic learners do their best work. As the way in shifts towards written prompts, that needs to keep being true rather than quietly stop being true.
Creative coding puts something on the screen immediately. Progress does not depend on getting through a wall of text first, which makes it one of the gentler doors into STEM for a learner who reads slowly.
Learners and teachers show us where the gap actually is, and it is usually in getting the question into the machine rather than in reading what comes back. That is the most useful signal we get, and it shapes what we research and what we train for.
Where the research happens, and where it reaches people.
p5js.ai is a research instrument as much as a product. It is free so that the people whose experience we most need to understand are not the ones priced out of it, and what we learn there about how dyslexic users write to a model goes straight back into what we train and build.
We are open to research collaborations, partnerships, and beta testing. Send us a note and a person will read it.