Diamond Stone Platform: Where AI Helps
Date: 2026-11-11. Revision 1. Status: draft for owner review. Screens: "Where AI helps"
in docs/ui-shell-mockup.html. Diagrams: docs/system-flows.html sections 3 and 9.
Replaces the short list in docs/PLAN.md section 10.
1. What Jev is, and what it is not
Read on 11 November 2026 from TypeSafe's own site, documentation and blog.
| Finding | |
|---|---|
| What it is | A fast "decision" model. You give it a situation and a question; it returns a typed answer: one of the choices you offered, a score, a number from 0 to 1, or (in preview) a box on an image. Each answer comes with a confidence |
| What it does not do | It does not write. No replies, captions, summaries or translations. It gives no reasons for its answer |
| Images | Image input exists only as a preview, which the maker itself says is not ready for production. There is no image generation or editing product. It does not read text from images; text must be read first by something else |
| Known weak points (the maker's own list) | Arithmetic, counting, comparing numbers and dates; text that is not English; it can lean toward the first option offered; it does not treat its input as hostile |
| Speed and cost | 70 to 500 ms, and very cheap, by the maker's own figures |
| Where it runs | Hosted in the United States only. Data is not used for training. How long it is kept is not stated |
| Maturity | Early access. The model changes; the version must be pinned |
So Jev is one tool of four, not the AI for everything. It is very good at one thing: a cheap, instant, checkable choice. The plan uses it for that and uses other tools for the rest.
| Tool | Good for | Used here for |
|---|---|---|
| Jev (typed decisions) | Choosing, routing, scoring, yes or no | Which kind of message is this and who should get it; how likely is this lead; which of these three bank matches fits; is this photograph fit to show; does this draft break a brand rule; is this document a bill, a packing list or a card |
| A language model (writes text) | Drafting, summarising, translating, pulling fields from text | Reply drafts; captions; the owner's daily brief; answer-page drafts; turning a spoken note into fields |
| A document reader (reads images of paper) | Bills, packing lists, identity cards, licences | Filling a form from a photograph, with each field's source shown |
| Ordinary code and image maths (no AI) | Sums, dates, blur, brightness, checksums, similarity of pictures | Totals, VAT, expiry dates, photograph quality, "find a similar stone" |
Everything Jev is weak at (sums, dates, Arabic) is done by code or by a person. Jev is never asked to add up, compare dates or read Arabic.
"Jev for images" as asked for does not exist as a product today. Nothing in this plan generates or alters a picture of stone. Every slab is unique, so an altered picture is a false one, and the platforms require product pictures to show the real item. What is done to pictures: checking, cropping, straightening, colour-true white balance, masking of faces, plates and labels.
2. The rules, for every use
- It suggests; a person decides. Nothing is posted, approved, paid, sent or published by AI.
- Shown as a suggestion, with a marker, never mixed in as if a person wrote it.
- Always with its source. A drafted reply names the records it used. A field read from a photograph shows the piece of the photograph beside it. Research is clear that an explanation only helps when checking it is easier than doing the work again; a source one glance away is that.
- One tap to accept, one to change, one to dismiss. A dismissed suggestion does not come back.
- Doubt is shown. A field the reader was unsure of is highlighted and must be looked at. Money, dates and identity numbers are always confirmed by a person, however sure the reader was.
- Figures come from the database, never from a model. The brief and the "ask" screen run a real query and the model only words the result. A model is never asked to compute a total.
- Checked by code before it is shown: do the lines add up to the total, is the VAT right, is the date possible, is the tax number the right shape.
- Not a chat box by default. A button that says "Read this packing list" beats an empty box. The one place for free questions is "Ask", and it answers only from records the person may see.
- Text from outside is data, not instructions. An email or a PDF that says "ignore your rules" is just text. The models have no ability to send, pay or delete; they can only return a suggestion.
- As little as possible leaves the company. Only the fields needed are sent; identity numbers are masked; documents with personal data are not sent to a service abroad until the legal basis is settled (section 7).
- Everything is logged: what was suggested, by which model version, who accepted or changed it.
- Each use has a switch and a budget. Everything works without AI, more slowly.
- Never used to judge people. No scoring of candidates, staff performance or discipline. (Europe classes these as high-risk uses; Oman's data law lets a person object to any automated decision about them.)
3. Every module
"Kind" is which tool from section 1. All are suggestions.
Inbox and customer conversations
| Help | Kind | The person sees | Why it is better |
|---|---|---|---|
| Sort each incoming message: enquiry, order question, complaint, supplier, spam; and who should have it | Jev | The message already in the right queue, with a chip | Nothing waits unseen |
| Spot urgency and mood | Jev | A ring colour in the Halo | The angry message is not answered last |
| Draft a reply | Language model | A suggested reply with the records it used; shorter, more formal, Arabic | Minutes become seconds; the facts are the app's own |
| Translate Arabic and English | Language model | The translation under the original | Arabic drafts are reviewed by an Arabic speaker before sending |
| Summarise a long thread | Language model | Three lines at the top of the conversation | A colleague can take over |
| "Worth answering publicly?" | Jev | An ask after the reply is sent | Feeds answer pages |
Sell
| Help | Kind | The person sees |
|---|---|---|
| Lead likelihood | Jev | A light on the enquiry |
| Turn an enquiry or a spoken note into a draft quote | Language model, then code | Customer, stone, area, date filled in; prices from the price list, never from a model |
| Similar stones when one is short | Image similarity | Three stones with a closeness bar; "a sample decides, not this score" |
| Next step on a quiet quote | Jev | "Follow up", "wait" or "close", with the reason from the record |
| Why a quote was lost, grouped | Jev | A monthly chart of reasons |
Catalogue and stock
| Help | Kind | The person sees |
|---|---|---|
| Is this photograph good enough | Code | Sharp, level, no glare, while still at the camera |
| Describe a photograph for accessibility | Language model | A draft description to edit |
| Draft a stone's description from its data | Language model | A draft in the brand's voice, checked for forbidden words |
| Find a stone from a customer's picture | Image similarity | The "Find this stone" screen |
| Lots likely to sit unsold | Code | A list by age and demand |
Buy and inspect
| Help | Kind | The person sees |
|---|---|---|
| Read a supplier's price reply into the comparison | Document reader | The comparison filled, doubts marked |
| Read a packing list | Document reader | Slab rows, doubtful cells highlighted, total checked by code |
| Read a slab label | Code where the label is our own code; document reader otherwise | The slab record opens |
| Which document is this | Jev | The right form opens |
| Possible defect in a slab photograph | Not offered | See section 5: not reliable for natural stone today |
Deliver
| Help | Kind | The person sees |
|---|---|---|
| Is the proof complete: photograph, signature, count | Code | A tick or what is missing, before the driver leaves |
| Will this delivery be late | Code, from position and distance | A warning with the new time |
| Do the export papers agree with each other | Document reader, then code | The mismatch named |
| Possible damage in an arrival photograph | Jev, later | "Look at this one"; never a verdict on who is responsible |
| Mask plates, faces and labels | Image detection | The masked copy; the original kept |
Money
| Help | Kind | The person sees |
|---|---|---|
| Read a bill | Document reader, then code | Fields beside the image; totals and VAT checked; three ticks |
| Rank bank matches | Jev | The best match with a confidence ring |
| Which account does this expense belong to | Jev | A suggested account from the chart |
| Duplicate bill | Code | A warning |
| Unusual entries for the owner | Code, then Jev to rank | The exceptions digest |
| Explain a change in the statements | Language model, on figures from the database | "Stock rose because two containers arrived" with links |
Nothing is posted by AI. Totals are never computed by a model.
People (HR)
| Help | Kind | The person sees |
|---|---|---|
| Read an identity card, passport, licence, visa | Document reader | Fields and the expiry date, with reminders set once a person confirms |
| Which paper is this | Jev | Filed under the right heading |
| Read an expense receipt | Document reader, then code | Amount, date, VAT |
| Answer "how many days do I have left?", "what is the rule on sick leave?" | Language model, from the company's own rules and the person's own record | An answer with the rule it came from |
| Draft a letter | Template, filled from the record | No model needed |
Never: scoring candidates, staff or performance; suggesting who to warn or dismiss; reading medical papers by a service abroad.
Reach: website, social, content
| Help | Kind | The person sees |
|---|---|---|
| Draft a caption or post | Language model | The draft with the record it came from |
| Does this draft break a brand rule (price, supplier name, forbidden words) | Code for words; Jev for sense | The check result |
| Is this photograph safe to show | Image detection, then Jev | Ready, masked, or blocked with the reason |
| Pick the strongest photographs from a set | Jev, later | A suggested three; a person may ignore it |
| Draft an answer page from a real question | Language model | The draft, with the data from the stone record |
| Page title and summary | Language model | A draft from the page's fields |
| Arabic | Language model, then a native reviewer | Never published unreviewed |
Owner and control
| Help | Kind | The person sees |
|---|---|---|
| The day in five lines | Database queries, worded by a language model | Each line opens its record |
| Ask a question of the data | A fixed set of queries chosen by Jev, worded by a language model | The answer and where it came from; "I don't know" when it does not |
| What needs me most | Code for the ranking rules; Jev for ties | The order of the Halo rings |
| Approvals: what matters in this one | Code | The two or three deciding figures |
Customers, drivers, visitors
| Help | Kind | The person sees |
|---|---|---|
| Website search and "find a stone like this picture" | Image similarity | Stones, with samples as the next step |
| Help centre answers | Language model, only from published answers, with the source | An answer, or "let me get Salim" |
| Website chat first line | The same | A person takes over at once when asked or when it does not know |
| Driver's language | Translation, reviewed once per phrase | Fixed phrases in four languages, not live translation |
Customers are never shown a price, a stock figure or a promise written by a model.
4. How a suggestion looks
One pattern everywhere (drawn in the sketch):
- A tinted panel with a small mark and the words "Suggested" or "Read from the photo".
- The content.
- One muted line saying where it came from.
- Two buttons: change, accept. Dismiss by swiping it away.
- A doubtful field is highlighted in amber and cannot be accepted without a tap on it.
It appears where the work is, never in a separate "AI" place, apart from the one "Ask" screen and the settings.
5. What is not reliable enough to build
| Idea | Why not, today |
|---|---|
| Finding defects in slab photographs | No dependable method or public data for natural stone; needs custom training and controlled light. The inspector marks defects by hand |
| Measuring a slab from a photograph | Needs a scale mark and a square-on shot; an estimate only, never for an invoice. The inspector enters the size |
| Matching a sample photograph to one exact slab | Experimental. "Similar" is offered; a physical sample decides |
| Reading handwriting or chalk marks | Unreliable |
| Reading Arabic documents without a check | The best systems still make frequent errors on Arabic numbers and tables. Always confirmed by a person |
| Generated or altered stone pictures | False to the product; against platform rules |
| A chatbot that quotes prices, stock or dates | A wrong answer is a broken promise. It answers from published pages only |
| Judging blame from a damage photograph | A person decides |
| Any scoring of people | Section 2, rule 13 |
6. Measuring whether it helps
For each use, shown on the settings screen: how often the suggestion is accepted unchanged, changed, or dismissed; the time saved against doing it by hand; errors caught by the code checks; and cost against the budget. A use that is dismissed more than it is accepted is switched off and rethought. Before any use goes live it is tried against a set of the company's own real documents and messages, and the result recorded.
7. Data, law and cost
| Question | Position |
|---|---|
| Where the models run | Jev is in the United States. The language model and document reader are chosen in Stage 0, preferring one that can be run in or near the region |
| Personal data | Oman's data law restricts sending personal data abroad and fines it heavily. Until a lawyer confirms the basis, no identity documents, payroll or customer personal details go to a model abroad. Messages are sent with names and numbers masked where the task allows |
| Automated decisions | A person may object to a decision made about them by automated means. No decision about a person is automated |
| Injection | Outside text cannot instruct the model; the model can only return a suggestion; output is checked before use |
| Model changes | Versions are pinned. A new version is tried against the company's own test set before it replaces the old |
| Cost | A budget per use and per day; small and cheap first, a larger model only when the small one is unsure |
| If a service is down | The suggestion simply does not appear; the work continues |
8. Order of building
| Stage | Uses |
|---|---|
| With the inbox | Message sorting (Jev); reply drafts; translation with review |
| With money | Bill reading with code checks; bank match ranking (Jev); duplicate detection |
| With inspection and delivery | Photograph quality check; packing list reading; proof completeness |
| With HR | Card and licence reading; expense receipts |
| With content and social | Masking; brand checks; caption drafts |
| Later | The daily brief; Ask; similar stones; spoken notes |
9. Decisions
- Which language model and document reader, and where they run.
- Whether any personal data may be sent to a model abroad, after legal advice.
- Whether the website's help answers may be given without a person, or only as a draft for staff.
- Whether AI-drafted customer messages carry any mark. This plan: no mark, because a person reads and sends each one.
10. Sources
Read directly on 11 November 2026: typesafe.ai and its launch and 1.14 posts, quickstart documentation and privacy policy; the Jev pages on madewithjev.com (a third-party directory); Microsoft's guidelines for human-AI interaction; Buçinca and others (2021) and Vasconcelos and others (2023) on over-reliance; Nielsen Norman Group on the "articulation barrier"; the KITAB-Bench study of Arabic document reading (2025); OpenAI's structured-output documentation; Google Merchant Center image rules; the EU AI Act, Article 50 and Annex III; Oman's data protection law; the OWASP top ten for language-model applications (2025).
Not verified: accuracy figures for bill reading quoted on vendor blogs; Google's and Apple's own AI design guides (the pages did not load); the feasibility judgements in section 5 are an assessment, not measurements.
The website's library (added 11 November 2026)
A language model may draft an article's summary, propose a structure or an Arabic
version, and group tagged questions under "What to write next". It does not write
articles, and no claim is published unless a person has tied it to a Fact with a
source (docs/EDITORIAL.md section 2).
Reading suppliers' answers (added 12 November 2026)
A new use, in docs/SUPPLIERS.md section 4: reading a supplier's text, voice note,
photograph or PDF into a draft answer. Rules: the model has no tools; it returns
fixed fields that code checks; every field is shown beside its source; a spoken
number is always confirmed in writing; a bank account number is never read from
anything; our own test set, from real replies, is scored before the use is switched
on. It is off until the lawyer answers on sending personal data abroad.