Practice
We take problems that already cost hours every week, and we leave them with a score the client can keep reading after we go.
Everydayintelligence is an applied AI studio founded in 2026 in Singapore. The practice is organised around six kinds of work that sit inside ordinary operations: finding things in the firm’s own files, moving fields out of documents, putting a bounded assistant next to a desk, offering a hint at the moment a person has to choose, and the evaluation that decides whether any of that is allowed to run. Discovery is the sixth area, and it comes first, because a polished system on the wrong task is still a waste of a quarter.
Each area below names the artefact you would actually receive, the first two weeks of work, and the conditions under which we would decline. If your problem sits across two areas, we still start with one, so the evaluation set can be scored without argument about what “done” means.
Two weeks beside the team. We time the operations, read a sample of the files, and write down the questions people already ask each other. The artefact is a short brief: one or two tasks where a system would change the week, the tasks we would leave alone, a first list of evaluation questions, and a constraint note on data residency. We have declined work at the end of discovery when the hours were real and the files could not support an answer. That is a successful sprint.
Search and answer over contracts, standard operating procedures, tickets, and manuals that already belong to the firm. We store embeddings, which are numerical summaries of passages so that similar meaning sits near similar meaning, and we require every answer to cite a source. The evaluation set is collected from real questions before a wider group sees a box they can type into. A fluent answer without a citation is treated as a miss.
Extraction of fields from invoices, bills of lading, claim forms, and delivery notes, written into the system of record the team already uses. Uncertain readings become a queue with the original image beside the guess. We agree the confidence line in writing, so a smeared vessel name is a human task, and a clean invoice number is a posting. Volume in the low thousands per month is a typical shape for the desks we sit with.
Hints at the moment a person has to choose: which claim to pick from a queue, which shipment to hold, which contract clause to read next. The system proposes; a person decides. We keep a record of the hint, the features it used, and the action taken, so a supervisor can see whether the hint was used and whether it aged well. Each row is a ranked list with reasons a colleague could argue with.
A bounded colleague inside Slack, a portal, or a CRM, with the same access rights as the person it helps, and a journal of every reply. It may refuse a question outside its corpus. We design the refusal text as carefully as the answer, because a night-desk operator who receives a guess will act on it. Typical first step: we watch a shift and list the questions that repeat after the office has gone home.
The test set, the reruns when a model or a folder changes, the dashboard, and the training of the person who will own those scores. Drift is the name we give to a slow slide in quality as files, models, or habits move. We treat evaluation as a product in its own right, because a system without a score becomes a story, and a story cannot be handed to a new operator in six months’ time.
How we start
The first fortnight is the same shape even when the later build is not.
We ask for a named counterpart who can walk us through the desk, and for access to a sample of the files under a short confidentiality note. A cleaned warehouse is unnecessary at this stage. We need enough of the mess to see which questions the files can actually answer. We sit with the people who do the work and with the people who sponsor it, and we copy the last questions they typed, spelling and all.
At the end of two weeks you receive a brief of a few pages, a first evaluation list, and a recommendation to continue, to narrow, or to stop. If we recommend a stop, we still invoice the sprint. The value of discovery is a decision made with evidence, including the decision to spend the next quarter on something else.
Indicative fees
Singapore, 2026. Scoped after discovery. Figures are starting points for a single desk. A firm-wide programme is a separate conversation.
Fees below are indicative for a first engagement in our first year. They assume one corpus, one interface, and one owner on your side. Travel inside Singapore is included. Hosted-model usage is billed through, at cost, once the prototype is live. We will not quote a hardening phase until the evaluation set has been scored at least once.
| Shape | Duration | Indicative fee |
|---|---|---|
| Discovery sprint | 2 weeks | from S$14,000 |
| Working prototype | 3 to 5 weeks | from S$36,000 |
| Hardening | 4 to 8 weeks | from S$48,000 |
| Handover and care | monthly | from S$4,800 / month |
A typical first conversation is a discovery sprint. If the sprint says the files cannot support the task, we stop there. If it says they can, the prototype quote follows from the brief.
Boundaries
A short list of work we do not take in our first year.
We decline consumer-facing chat as a product in itself, open-ended “ask the company anything” boxes with no corpus boundary, and work that has no named operator on the client side. We also decline medical diagnosis, and any system whose purpose is to replace a licensed professional’s sign-off. If the brief is a slide deck about transformation with no desk attached, we will ask you to name the desk or we will pass.
We are a small studio in our first year. We take a handful of engagements at a time so that the people who write the evaluation set are still the people who read the monthly misses. If we are full, we will say so and offer a later window.
Next
A useful brief names the operation, the systems already in use, and whether the documents may leave Singapore. We reply within one working day.