Two films, four shorts, one platform

All six are real renders of the product's own chrome, animated frame by frame in Remotion from the same data the application holds. The features take one operation end to end at opposite ends of the scale. Each short holds a single moment still: what one changed field touches, what happens after a human overrules the machine, who approves when nobody is at a desk, and what it costs to correct a model that turned out wrong.

2:50 · with sound

Blue Valleythe CEO's cut

A 140-person B2B software company plugs in seven systems at once — email, Slack, Jira, Stripe, Calendar, Meet, Docs — and gets one model back. 24 310 records, 90 days of history, one morning briefing that joins a failed payment, a blocked sprint, a sev2 and a QBR into a single account.

seven connectorsdrafted ontologythe briefingthe pulsethe SDK

2:23 · with sound

After 8the owner-operator cut

The same platform from the other end of the scale: one person, one mailbox, an event production agency. The inbox is read, the operation is described back, the schema is drafted, and a summer party for 120 guests gets produced — every write a proposal the owner approves.

one mailboxthe analysisthe schemavendors & offerthe week

0:58 · with sound

Impactshort · one date moves

A wedding shifts from 12 June to the 19th — one field, on one record. But a venue was held, two crew were confirmed, an invoice was dated, and three email threads still carry the old date. The impact reviewer follows the links, reads every one of them, and comes back with seven proposals. She approves six and rings the DJ herself.

one field movesthe reviewer readsseven proposalssix approved

1:08 · with sound

Lessonsshort · it gets one wrong

The inbox orchestrator drafts a reply promising the fix ships Friday. It is wrong — the ticket is blocked and Friday is not real. She does not just reject it: she edits it, files her version as her own action, and the reason lands in the audit trail. Four of five drafts from that agent were rejected; the pattern gets mined, written down and confirmed. Thursday, same situation — and the draft no longer promises a date it cannot keep.

a wrong draftedit & approvethe pattern, minedcalibrated

1:05 · with sound

Pocketshort · nobody is at a desk

An independent garage — two ramps, no office staff. At 09:14 a customer emails about a grinding noise; triage files three proposals into an empty office. They go to the owner's pocket instead: a bot paired once with a code, /pending, three cards, two taps and a dictated reply. By the time he washes his hands it is all in the trail — the same verbs, the same queue, the same audit. Not a lite version.

triage, unattendedthree cardsapproved with a thumbsame pipeline

1:02 · with sound

You'll get it wrongshort · correcting a live model

The objection none of the other films answer. Six months in, deposit terms have been going into a notes field three different ways — unsearchable and unenforceable. So the model changes: two fields, an enum and a rule that watches them, proposed on a live workspace with 340 bookings already in it. Approve, and the form, the list, the board filter and what the assistant may read all reshape with it. No migration, no deploy, no ticket.

a hole in the modelfour approvalseverything reshapesno deploy

Rendered with narration and score, and served at 1080p — with a 4K switch in the player wherever a 4K master exists. Everything below runs live in this page: the same scenes, re-cut for the browser.

Systems that never talk to each other

Money in one system, delivery in another, the decision in a third, and the outside world in a mailbox. Nothing links an at-risk renewal to the sprint that slipped, to the meeting where it was raised. Connect the sources, and the platform reads all of them before it proposes anything.

Figures from the Blue Valley film: seven connectors, ninety days of history, one model. Email, documents and Telegram ship in the box today — the rest arrive as packages against the same manifest, which is the whole point of the contract below.

The ontology is data, not code

Object types, their fields, their links — rows you create and reshape at runtime. There is deliberately no reference field: one relationship mechanism means one place to look for the graph and one backlinks panel that is always complete. Describe your operation in plain language and the ontology assistant drafts the schema — as proposals you approve inline. It cannot touch the schema on its own.

One write path, for people and machines alike

Reading is free — every assistant tool is read-only. Writing is not. Every mutation, from any source, is a verb going through one pipeline. The only difference between you and an agent is the mode.

The actions table does double duty on purpose: the approval queue while a proposal is pending, and the audit trail forever after.

One question, six sources, four proposals

The assistant looks the records up with read-only tools, shows what it found and where each piece came from, and files what it wants to do as approvable cards.

Every write, whoever made it

Connectors, rules, assistants and you, on the same ledger — with the full trace of every AI turn behind it: tool calls, reasoning, errors.

Every capability is a package

The kernel owns the data and the single write pipeline. Packages bring everything else — each one a manifest declaring its settings, traits, governed verbs, agents, job handlers, hooks and read tools. At boot they register into the kernel registries. The kernel never names a package.

No package imports another. They meet in the kernel registries and in the actions table — enforced by boundaries.test.ts over every import form.

Where you meet it

Everything is connected: reminders live on records, on the dashboard and on the calendar; proposals link to the records they touch and can be approved from Telegram; email drafts know the correspondent's schedule; inbound attachments save onto records and record documents attach to outbound mail.

A single process you can actually reason about

Next.js App Router over SQLite in WAL mode with a single writer, one durable job queue, one LLM gateway. Deploy-per-client, one tenant per process. The whole workspace is one database file — exportable and restorable from Settings.

Held by a contract battery — package boundaries, table ownership, registry shape and the kernel's own invariants — that gates every release.