01
Start with a real constraint
A game must feel good in somebody’s hand. A garden plan must agree with the ground. The work begins where a generic AI answer stops being enough.
Our live AI projects
We do not make AI demos. We build things that have to work.
A spatial game whose results are replayed by the server. A real garden where every AI idea must survive measurements, money, weather, and human judgement. These are the laboratories behind Collab365’s approach to AI.
Collab365 / AI lab
Live
Foldami
Deterministic paper mechanics, verified at the edge.
In progress
AI Garden Project
AI ideas tested against a real UK garden.
01Hypothesis
02System
03Verification
04Ship
What “deep” means to us
Depth is what happens after the first prompt.
AI gives a small team extraordinary reach. It also creates plausible mistakes at extraordinary speed. Our work is about building the systems that separate those two things.
02
Turn judgement into a system
We give sources, states, rules, tests, and acceptance criteria a permanent home so quality does not depend on remembering one clever conversation.
03
Make verification structural
Foldami replays the move. The garden traces the claim. In both projects, the check is part of the architecture rather than a final polish pass.
04
Put it in front of reality
We ship the game and publish the garden record. The interesting evidence is what survives contact with players, measurements, budgets, weather, and people.
01 · Live at foldami.com
Product experiment
A new puzzle verb, built into a complete product system.
Foldami begins with one simple instruction: fold the paper until nothing is left on it. Underneath that quiet surface is a deterministic engine, a generated puzzle catalogue, a touch-first product, an embeddable package, and a Worker that refuses to believe the client’s score.
Pure game engineServer replayPWAClosed-shadow embed
foldami.com
ENGINE
Same rules, every replay
VERIFY
Worker derives the result
DELIVER
Edge app + embed package
Inside the system
The parts a screenshot cannot show
The puzzle is the visible layer. These are the systems that make it repeatable, portable, and trustworthy.
A deterministic paper engine
Every fold is a repeatable transformation of a layered sheet. The engine has no DOM, no clock, and no random state hidden inside it.
Seeded puzzles · pure reducer · replayable moves
The feel is engineered too
Touch gestures, fold rehearsal, paper motion, audio, haptics, progression, sharing, and accessibility are treated as product systems—not decorative extras.
One-thumb play · daily sheets · quiet mode
The client cannot invent a score
A signed, one-use session names the puzzle. The Worker regenerates the board, replays the submitted folds, and derives the result for itself.
No submitted score · no trusted board · no reusable nonce
Built for the edge
Expensive puzzle generation happens offline. The live Worker serves the PWA, verifies linear replays, records results, and keeps the hot path deliberately small.
Cloudflare Worker · D1/KV · offline catalogue
A product that can travel
A separate embed package can mount the game into a closed shadow root. Host-page isolation is useful; score integrity still comes from server replay.
Script-tag embed · scoped styles · verified outcomes
AI helped us preserve the hard-won parts
Instead of repeatedly rewriting the clever-looking code, we isolated the tested engine and rebuilt the risky edges around it with explicit contracts.
Measured failures · regression tests · documented decisions
AI proposal · concept visual
Truth boundary
Beautiful is not the same as true.
Every concept keeps its status until measurement, professional input, and the build earn something stronger.
02 · Building in public
Real-world experiment
An AI design experiment that reality is allowed to prove wrong.
The AI Garden Project asks whether two ordinary UK homeowners can use AI, professional input, and their own judgement to redesign and build a real garden without making expensive mistakes. The public record includes the attractive ideas, the broken geometry, the checks, the decisions, and what the physical result eventually changes.
Current truth
The physical baseline is still in progress. Existing visuals and plans are clearly labelled as concepts or working material—not construction-ready proof.
Inside the system
The machinery behind the public story
This is not a feed of AI renders. It is a governed record of how inputs, decisions, evidence, and outcomes change.
Ground truth before prompting
Hand measurements, a zoned plan, a photo atlas, fixed features, and repeatable viewpoints give the models something more useful than a beautiful guess.
Measured inputs · photo atlas · fixed reference facts
Every claim leaves a trail
Sources, claims, decisions, assumptions, risks, and changes live in linked registers. A confident AI sentence never promotes itself into a verified fact.
Source register · evidence ledger · decision log
One record becomes many useful views
Validated source records generate the Journal, Problems, Outcomes, plan, sitemap, and typed public libraries. A broken record fails the build on purpose.
Content generators · schema gates · no draft leakage
Concept and evidence stay different
AI visuals are dated and labelled as proposals. Public photographs, working plans, edited images, and future as-built evidence each carry their real status.
Before · proposed · built · drift
Reality gets the deciding vote
The garden is the test bench. Measurements, levels, drainage, quotations, build discoveries, planting, and professional checks can overturn the model.
Physical build · accountable checks · visible uncertainty
Publishing is governed, not improvised
Privacy, provenance, claim traceability, public review, structured data, and a clear free-versus-paid boundary are designed into the publishing workflow.
Human review · public-safe exports · machine-readable proof
What comes back into Collab365
Two different projects. The same hard lesson.
AI can let a small team attempt work that once needed far more people. But speed only becomes an advantage when the sources, state, rules, checks, and owner are explicit. That is the principle we carry into Spaces, Blueprints, and every system we build for customers.
Read the full Collab365 rebuild storyVerification belongs in the architecture.
Evidence and uncertainty stay attached to the claim.
Repeatable workflows beat one brilliant prompt.
Human judgement owns the outcome.