Futureproof, the AI Garden Project, Foldami, AppStore2031, and the small amount of client work that keeps the research honest.
Experiments and client work · 01
Why does Collab365 build side projects?
Collab365 builds side projects to learn AI by building with it, in public, against problems that punish a wrong answer. A game has to feel right in your hand, a garden plan has to agree with the ground, a forecast has to survive being challenged. The methods that hold up come back into Spaces.
Reading about AI teaches you very little. Shipping something that real people use, where a confident but wrong answer is immediately obvious, teaches you a great deal.
Each project is chosen because it has a hard constraint that generic AI output cannot fake its way past. That is the point of them.
To be clear: These are not client demos and not products Collab365 is trying to sell you.
Do the side projects distract from building Spaces?
The side projects are deliberately small, and each one earns its place by teaching something Collab365 Spaces needs: verification that cannot be faked, evidence that survives a challenge, publishing that stays honest at volume. When a project stops teaching, it stops getting attention.
The honest version is that a two-person company cannot afford hobbies. What it also cannot afford is learning AI from other people’s blog posts while the ground moves.
The architecture behind Spaces came out of exactly this kind of building.
Experiments and client work · 03
What is Collab365 Futureproof?
Futureproof is a free research tool that scores US and UK occupations task by task against official government data, so you can see which parts of your job are genuinely exposed to AI rather than guessing from a headline. No signup, and the data is downloadable.
It works at the level of tasks, not job titles, which is the part most coverage gets wrong. Very few jobs disappear whole. Parts of them change, and the parts differ enormously between two people with the same title.
The rankings show both ends together with the uncertainty visible, the method page sets out the rubric and where it falls short, and every score is downloadable as open CSV under CC BY 4.0.
It lives at futureproof.collab365.com.
To be clear: Futureproof is not a prediction that your job is safe or doomed, and it is not a lead magnet with a form in front of it.
The AI Garden Project is a real garden being designed and built with AI, where the ground is allowed to prove the AI wrong. Measurements, levels, drainage, quotes and professional checks can overturn the model, and when they do, the record says so rather than quietly updating.
It exists because a garden is unusually honest. You cannot argue a slope into being flat, and a plant either survives or it does not.
The method is the interesting part. Hand measurements and a photo atlas go in before any prompting. Sources, claims, decisions, assumptions and risks live in linked registers, so a confident AI sentence never promotes itself into a verified fact. AI visuals are dated and labelled as proposals, kept separate from photographs of what was actually built.
That separation of proposal from evidence is the single most useful idea it has produced, and it now shapes how Collab365 publishes research.
To be clear: It is not a gardening service and not a showcase of pretty AI renders.
Foldami is a paper folding puzzle game, built to test whether AI could help ship a complete product rather than a demo. Every fold is a repeatable transformation, and the server replays your moves to derive the result, so the game cannot be talked into believing a score.
Games are a hard test because feel cannot be faked. Touch, motion, audio, haptics and progression either work in somebody’s hand or they do not, and no amount of confident explanation fixes it.
The lesson that carried across was structural verification: build the check into the architecture rather than trusting what you are told.
AppStore2031 is a contestable forecast of the products and the kinds of work AI could create by 2031. Every entry keeps its source, its assumptions and the challenges made against it attached, and any forecast can be publicly withdrawn when the evidence turns against it rather than quietly edited.
Most predictions about AI are unfalsifiable, which is why they age so badly and cost their authors nothing.
The design question here was what a forecast looks like when being wrong has to be visible. The answer was to make withdrawal a first-class outcome rather than an embarrassment.
Do the experiments cost anything, and do I need a Spaces account?
They are free and open, and none of them require a Collab365 account. Futureproof asks for no signup and publishes its data openly, the garden record is public, and the forecasts are readable by anyone. They exist to show the working, not to collect email addresses.
If something in them is useful to you and you never touch Collab365 Spaces, that is a perfectly good outcome.
Experiments and client work · 08
Does Collab365 do client work?
Collab365 takes on a small amount of client work, and it keeps the research honest by putting it in front of real businesses. Priorslee Motor Spares is the published example: a local car parts firm whose outdated site was rebuilt into something useful, and put live the same working day.
That timing is a fact about one project, not a promise about yours. It went that fast because the scope was clear and the owner could make decisions on the spot.
If you are wondering whether something similar would work for your business, ask through the contact page and you will get a straight answer, including when the honest answer is no.