Ateliersavant In-House
LOOM — The Relationship OS That Refuses to Look Like It Works
A founder carries a hundred threads and can hold six. LOOM holds all of them — every venture, every conversation, every person and how they came into your life — and hands them back the moment they matter. Not a database you query. Something that knows the whole shape of what you are building, and talks to you about it like someone who has been there the whole time.
WHAT IT IS
LOOM weaves the threads. Every venture, every negotiation, every person and the room you met them in — held as one fabric, read at a glance, and answered aloud.
You do not open it and search. You ask it something, in your own voice, in the middle of doing something else, and it answers with the whole context of your life's work behind the answer — because it has been holding that context the entire time. It knows which negotiation has gone quiet. It knows who introduced you to the person you are about to call, and when, and what came of it. It knows that the venture in Osaka and the supplier in Reims are the same problem wearing different clothes.
The Category Builds Silos. The Work Lives Between Them.
A founder running a portfolio of ventures across several jurisdictions does not have a data problem. They have a fabric problem. The files are in one place, the notes in another, the correspondence in a third, the relationships nowhere at all except in a head that is already full. Every product sold into this gap solves it by becoming another silo — one more system that wants to own the files, own the notes, own the calendar, and be opened every morning.
LOOM was built on the refusal of that model. It stores almost nothing. Box remains the file store, Obsidian the note vault, Gmail and Calendar the communications layer. LOOM holds pointers and intelligence, and weaves the threads — conversations, relationships, projects, entities — into one fabric read at a glance. It is connective tissue, not a destination.
LOOM — General Presentation. The system in operation.
The Membrane
Not a screen. A material — with a light source, a surface, and things moving underneath it.
Almost every AI product on the market wears the same face: a chat column, a sidebar, a blinking cursor, a gradient. The interface is a container for text and admits it. LOOM was built on a single physical premise instead — the page is vellum, what is printed on it is ink, and what moves beneath it deforms it, lit by one raking light from the upper left. Every element in the product either takes ink or takes relief. Nothing floats.
The consequence is a rule most design systems never state: the system may not invent a second visual vocabulary when it already has one. The product's central instrument is a cymatic plate — a real superposition of drifting wave sources, rendered live. When insight markers needed a figure, an ornament was designed for them, and then thrown away. They now carry the instrument's own figure, frozen: the same renderer with the clock held at a fixed seed, so each marker is that plate caught at a different instant. No two are identical, and nothing was invented.
The wordmark follows the same logic to its conclusion. It had been printed in ink with a bronze initial. It is now a blind deboss — struck into the vellum, taking no ink at all, upper-left walls in shadow, lower-right lips catching the light. A blind deboss takes no ink by definition; a coloured letter inside one is a contradiction. The bronze was retired.
Two failed attempts are on the record, because they are the proof. Inking the wave crests produced soft blobs. Inking the interference fringes produced a dark granular mark that read as a sacred pattern. Both were foreign objects sitting on the page rather than deformations of it. The accepted version writes the emboss as alpha rather than colour — so it lands tone on tone on whatever the page already is, and re-strikes itself when the theme inverts.
The membrane in motion — the instrument, the relief, and the strike.
Three Failures That Looked
Exactly Like Success.
Published in full, because the architecture only makes sense in their light. Each one produced output. Each one was invisible. Each one is now closed in code rather than in intent — because intent had already been correct, and had not been enough.
It answered a television for three hours.
With the owner abroad, the machine transcribed a science-fiction soundtrack and replied to it — spending model turns, polluting its own conversation history, and in one instance writing a record naming a real person.
The forensic finding is the interesting part. Not one answered line contained a wake word. A browser tab left open on the machine had engaged the listening loop unconditionally, and every line of dialogue renewed the window. Nothing about a forgotten tab says its owner is in the country.
Closed: the system now reads the machine's own idle time. Past thirty minutes unattended, a literal summons is required, there is no follow-up window because there is nobody to follow up, and at most six turns an hour may be spent on an empty room. Ten transcripts from that night are now permanent regression fixtures.
A week of speech produced one garbled sentence.
Mobile speech recognition failed almost totally in the field — words died before reaching the system. Worse, the owner had not realised the phone had ever activated capture at all.
The second finding outranked the first. A microphone whose state a person cannot perceive is not a feature with a bug; it is a different product than the one they agreed to.
Closed: recognition on mobile was deleted outright rather than improved — guessing badly is worse than not guessing. The microphone now opens only while a finger is held on it, with no code path that opens it otherwise, so the operating system's own recording indicator burns for exactly the length of the press. A warm stream between presses would have been faster and would have made the honest claim untrue; it was rejected. The change to capture behaviour forced every user to re-accept the terms.
It had been broken for three days, and said so in a British accent.
A stray credential in an environment file — never referenced by any line of the system's own code — silently redirected every request to metered billing. A spend cap then stopped the system dead. For three days every question failed, and every failure was dressed up as an answer: the provider's own complaint was filed as the system's reply, rendered on the page, and read aloud.
The mechanism is worth naming for anyone building on agent frameworks. An upstream refusal arrives marked as a successful turn whose text is the complaint. Every naive success check passes it straight through. The questions were marked answered, so nothing ever retried them.
Closed: the guarantee moved out of intent and into code. The child environment now strips those credentials by construction, so no stray line can re-create the leak, and the system states which credential it holds aloud at every boot. A refusal now throws rather than answers. A leak that hid for three days can hide for three months; the boot line is the smoke alarm.
Intelligence at
a fifth of a cent.
The constraint that shaped everything else was commercial, and deliberately brutal: a hard ceiling of ten dollars a month, with a target of zero to two. Not because the budget demanded it — but because a system that cannot justify its own running cost has not earned the right to be in the daily loop.
That figure is the whole argument. The category's assumption is that intelligence of this depth belongs to enterprises with enterprise budgets — that a system holding your entire portfolio, listening in the room, drafting in your voice and maintaining a live relationship graph is a seat licence, a platform fee, a procurement cycle. LOOM refutes it at roughly two-tenths of a cent per answer, and refutes it on infrastructure anyone can reach.
The discipline that gets you there is architectural, not frugal. Store pointers, not copies. Do not rebuild what Box, Obsidian and a mailbox already do — link to them. Keep the expensive brain on hardware you already own and let a small understudy cover it. Publish the cap, check it before spending, and say so aloud when it is reached. The result is a system whose running cost is small enough to be irrelevant to the decision to use it — which is the only cost structure that survives daily habit.
Authorisation Is the Database, Not the Code
LOOM is multi-user by design: the principal plus partners and contributors, each with per-project access. The security model refuses the usual arrangement in which application code decides who may see what. Every table is default-deny at the row level, and every conversational route executes under the calling user's own credentials, so the database — not the code path — is the entire authorisation model. Twenty-three access assertions are verified from the client, as owner, as read-partner, as edit-partner, as no-access, and as anonymous.
Two rules follow that are unusual enough to state plainly. First, correspondence is stored as summary, never as raw content: partners see a summarised negotiation timeline, never an inbox. Second, no partner ever draws on the principal's AI subscription — each conversational surface runs on its user's own key, resolved from an encrypted vault, so a trust boundary is also a billing boundary.
Nothing sends and nothing mutates the relationship graph autonomously. The system proposes — drafts, graph edges, context updates — into review queues, and a human validates before anything executes. A reply travels in-thread with a sixty-second window to recall it. The relationship graph is maintained by the system proposing edges from new correspondence and a person confirming with one tap; manual-only edge entry is treated as a failure mode, because a graph nobody maintains is a graph that lies.
Availability is handled by admission rather than aspiration. A warm agent session on the principal's own machine has right of first refusal on every question. When its pulse goes stale, a cloud understudy answers from a persisted snapshot — guarded three times against double-answering, capped at five dollars a month, checked before a single token is spent, and required to say so aloud when the cap is reached. A cap nobody can see is a cap nobody trusts.
One Act, End to End
Product Architecture
The system definition, the data model, the permission doctrine, and the decisions locked against relitigation before a line was written.
Interface & Identity
The membrane: a material premise, a single light source, a live cymatic instrument, and a wordmark struck rather than printed.
Voice Architecture
Capture, the wake decision, the presence gate, and what answers when the primary brain cannot.
Adversarial Review
Three lenses, twenty agents, findings independently verified. Fifteen confirmed and all closed the same night — including a prompt-injection path to code execution.
Why This Is a Break, Not an Increment
The entire consumer AI category is optimised to always produce an answer. Silence reads as failure, so systems are built never to be silent — and the result is a generation of products that degrade invisibly, answering confidently long after they have stopped being connected to anything true. Three times, in one month, LOOM did exactly that. The response was not to make it more reliable. It was to make failure structurally unable to disguise itself: a refusal that throws instead of answering, a credential named aloud at every boot, an unanswered question left recoverable rather than marked done, a cap that speaks when it is reached.
The interface argument is the same argument in another material. A product that claims to be honest and then wears an invented ornament has already broken the claim in the only place a user can see. Deriving every figure from the instrument, and striking the wordmark rather than printing it, is not decoration policy — it is the doctrine made visible at 46 pixels.
Ateliersavant builds its own systems on both sides of the curtain, which is why this record exists in this form. The failures are published because they are the asset: a firm that will tell you its product answered a television for three hours is a firm you can believe about the rest.
Next Case Study