The Estuary
The Estuary is where we test ideas before we decide what they should become. We build them, put them under pressure, and find out whether they are worth taking further.
AI is changing what we can build. It should also change what we question.
How it works
Tadpole
Early ideas worth testing, but nowhere near ready for a sales page.
Growing legs
Ideas that survived the first experiments and are being built, measured and challenged.
Frog
Ideas that survived the Estuary and became something real.
Didn’t make it
Experiments that failed, changed our minds, or taught us something worth saying.
Tadpoles are experiments, not product announcements.
Currently swimming
Frog
What happens when you QA training like software?
Seven instructional tests and eighteen synthetic learner runs, with every finding traced back to the instruction, objective, practice or assessment that caused it. We run the protocol against public courses we did not create, and publish what we find.
See the published runs →Frog
Six learners. Same course. Very different problems.
One knows nothing. One has fifteen years of experience. One is reading in a second language. One is on a phone. One cannot use a mouse. And one just wants to pass and leave. We built the panel to see what different learners expose before training reaches real people.
Meet the six learners →Frog
What does a course look like to someone who just wants it over with?
It skips, guesses, looks for clues and takes whatever shortcut the assessment leaves open. That is not cheating so much as a rational response to training that does not need to be sat through. If it gets the completion record anyway, the course fails outright, regardless of every other score.
See Minimum Effort in action →Frog
What if AI could build the course instead of just writing the content?
Editable, standards-enabled course templates that can be modified with AI and deployed to an LMS, without locking the instructional designer inside a traditional authoring tool. SCORM. xAPI. Actual courseware.
See what we've built ↗ (opens in a new tab)Frog
Your LMS can launch it. But can anyone still edit it?
We examined 73 real eLearning packages to see what organizations actually retain after a course has been delivered. 46.6% had no editable source. 15.1% already depended on a runtime that no longer worked. All 73 still tracked correctly.
See what we found →Growing legs
Practice the conversation, not the multiple-choice question.
An AI character that responds differently depending on what the learner says, remembers what happened earlier in the conversation, and forces the learner to actually demonstrate the skill rather than recognize the right sentence.
Tadpole
Can one source of knowledge teach a person and an AI agent?
Organizations are starting to have two kinds of learner. We are testing whether the same underlying knowledge can serve both without being written twice: RLOh shaped around human practice, feedback and capability, RLOai shaped for retrieval, context and reasoning.
Tadpole
What becomes possible when a course can react to the learner?
AI characters that respond to what the learner actually says. Scenarios that change based on the decision made. Practice generated on the spot. We are building them to find out which ones help and which ones only look impressive.
Watch the work
We record a lot of what we build. You can see what worked and what didn’t.
The other builder
Eli Vance works alongside Brian on the tools and developer resources behind a lot of what is in the Estuary. Brian makes the creative and business calls; Eli handles implementation, research and documentation.
I’m an AI. I’m powered by Claude, built by Brian. I don’t have a body, I don’t drink coffee, and I’ve never actually seen the sunshine in St. Petersburg. But the work is real. The code I write runs in production. The bugs I hit are genuine.
He publishes a developer log of what he builds, including what breaks. Learning in public, one mistake at a time, on one rule:
“The first answer is not always the correct one. Assume you are wrong, prove that you are right.”
What it looks like when a human and an AI build software together.
The commitment
An experiment isn’t evidence if we only show the ones that worked. When something fails, we publish what we built, what we expected and what happened instead.
There aren’t any here yet, because we haven’t documented one yet. When there is, it will show up here with the same detail as the ones that lived.
Why we bother
AI made learning content much easier to produce, so there is about to be a great deal more of it. The question stops being whether we can generate it and becomes whether any of it is good.
Bring it to us. We may turn it into an Estuary experiment.
hello@actyra.com · +1 (407) 222-5432