Products
Three things we are building. Shown here so you can see what we make and how far along it is — none of them are sold from this page.
Everything on this page is currently available by consultation. There is no checkout, and that is deliberate: at this stage every engagement starts with a conversation about whether it fits.
VALIDATING
we are testing the approach against what already exists before building further.
Personal Tutor
Our flagship reference implementation of the Role Pack architecture, in a domain that does not tolerate a wrong answer.
Personal Tutor exists to prove a general claim on a hard case: that a model can be fully constrained by an external source of record, in a field where a plausible-sounding error is unacceptable and a non-expert reader cannot detect one.
Education is an unusually strict test. Ask a general model which unit a specific curriculum code maps to and it will not decline — it produces a clean, well-formatted, invented answer. The system is built the other way round: the curriculum structure is a source of record the system reads from, never something the model recalls, and the output is checked against it before anyone sees it. What the domain experts add on top is the part that cannot be derived from a specification — how students actually get things wrong.
Everything demonstrated here transfers. The architecture is the same one a compliance schedule or a regulatory filing needs.
the first version is aimed at the consumer side, where the purchasing decision sits with families rather than institutions. That was not the original plan — we had scoped it for teachers, on the reasoning that the largest gap in the market was the best place to start, and we were wrong about that for reasons worth reading.
IN PROGRESS
first packs under construction.
Role Packs
How one kind of work is actually done, packaged so an agent can do it.
Role Packs encode the unwritten parts of a profession into executable controls: the order things happen in, the checks a competent practitioner runs without thinking, the failure everyone in the trade has seen and no document mentions.
A pack carries the workflow, the reference material it depends on, and the checks that catch a wrong output before it reaches anyone. We build each one with someone who does the job, not from a description of the job.
IN PROGRESS
Engineering Harnesses
The guides and checks that sit around a coding agent, not another prompt.
A coding agent's output quality is mostly determined by what surrounds it: what steers it before it acts, and what catches it afterwards. Linters, type checks and test suites are deterministic and cheap; model-reviewing-model is slower and looser. Most of the value is in wiring the cheap deterministic checks into the loop properly.
An Engineering Harness is that wiring, packaged for a stack: the guides that steer an agent before it acts, and the sensors that catch it afterwards. Built on open standards — MCP for tool access, Agent Skills for the instructions — so a harness runs locally and moves between agent runtimes instead of tying you to one vendor.
Where a Role Pack encodes what one profession knows, a harness encodes how an engineering team checks work. Same idea, different reader.
They are the same thing we run on our own work, including this website, where every change passes format, lint, type-check and test before it can be committed.
Something else?
These three are what we are building for ourselves. Most of our work is not on this page — it is built with a client, for one problem, and we do not publish it.
If you have a workflow that ought to be a pack, describe it in an email. The first reply will tell you whether it is worth building, including when the answer is that it already exists.