SignetAI
Lagos · Abuja · Alberta
§ Workbench · in plain terms

A private AI workbench that understands how your business works.

Not a chatbot bolted onto your files. It is a private workbench connected to your knowledge, workflows, records, media and approvals — so it can answer questions, write drafts, watch for problems, and handle routine follow-up. Always showing where its answers came from, and always with one of your people in charge.

§ 01

What it is, in one paragraph.

Think of it as a controlled workspace that understands your work, rules and evidence — and never forgets.

Your organisation’s knowledge — contracts, operating systems, regulator letters, board papers, site reports, policies, photos and project records — is connected and remembered. Ask a question in plain language and you get an answer in seconds, with the exact source it came from. Ask for a report and it drafts one, pulling the evidence straight from your own records.

It also keeps watch. Deadlines, targets, risks, a supplier going quiet — it flags them to the right person before they become a problem. And before anything leaves your business, one of your people reviews and approves it.

It’s where the work gets done — and where you can always prove what was done, when, and why.

The clever parts underneath are the same for every client. What changes is your sources, your rules, and what you ask it to do.

§ 02

Four things it always does — built in, not optional.

I

It shows its sources.

Every answer points to the exact source it came from, so you can open it and check for yourself. If it can’t find a solid answer in your approved knowledge, it says so — it doesn’t guess or make things up.

II

Your data stays yours.

Everything runs inside your own country, on systems you control. Your data and records are never sent abroad, and they are never used to train anyone else’s AI. Fully in line with Nigeria’s data-protection laws (NDPR and NDPA 2023), and the equivalent elsewhere.

III

Everything is recorded.

Every question, draft, approval and message is saved in a permanent record that can’t be quietly changed. If an auditor, a regulator, or your board ever asks what happened, you can show them exactly — down to the minute.

IV

A person always approves.

The AI does the work; one of your named staff approves before anything goes out. The AI is the tool, never the one who decides. Every outgoing item carries a real person’s name.

§ 03

Six things it does for you.

I

Answers questions

Ask anything about your business in plain language — “what’s the status of the Ijebu project?” — and get the answer in seconds, with the source attached.

Finds answers
II

Writes your reports

Board papers, donor reports, audit responses, briefings — drafted from your own records, every figure traceable, ready for someone to read and sign.

Drafts for review
III

Watches for problems

Keeps an eye on deadlines, targets and risks around the clock, and tells the right person — through the channel they actually check — before something slips.

Watches and alerts
IV

Checks your sites

Reads site photos, drone footage and CCTV like a trained inspector: safety gear not worn, asset corrosion or damage, construction progress, and perimeter incidents — flagged for your engineer to confirm.

Reads photos and video
V

Finds the right people

Spots the partners, suppliers or counterparties you should be talking to, and drafts the first message — ready to send from your own email.

Finds and contacts
VI

Keeps it all safe

The safety net underneath everything: nothing leaves without a person’s approval, and every action is recorded.

Governs everything
§ 04

The engineering underneath.

You don’t need to know any of this — but for the engineers and technical evaluators on your team, here is how the workbench actually finds and grounds its answers. It is the difference between a serious retrieval system and a chatbot pointed at a folder.

I

Hybrid retrieval.

Every question runs two searches at once — keyword (BM25) for the exact terms and semantic vector search for meaning — then merges them with reciprocal rank fusion. It finds the right passage whether your team recalls the exact wording or only the gist.

II

Contextual retrieval.

Before indexing, each passage is enriched with the context around it — which document, which section, which matter — so a lone figure in a table or a single clause in a 200-page contract is still found correctly on its own.

III

Grounded generation.

Answers and drafts are written only from the passages actually retrieved from your records, every statement tied to its source. Nothing is invented from the model’s general knowledge.

IV

It refuses rather than guesses.

If the evidence isn’t in your approved knowledge, the workbench says so plainly instead of producing a confident-sounding fabrication. Refusal is a feature, not a failure.

V

Private, in-country indexing.

Your documents are chunked, embedded and stored as a search index inside your own environment. The vector index and the text behind it never leave your control or cross a border.

VI

Built for messy, real records.

Scanned PDFs, tables, regulator letters, board packs, site photos and CCTV — handled together and re-indexed as new material is added, so answers stay current without a manual rebuild.

§ 05

Custom MCP connectors: how the workbench acts in your systems.

The workbench doesn’t just answer — it can act in your systems through custom MCP connectors.

An AI on its own can only talk. To be useful it has to reach your real systems — your document store, your calendar, your email, your finance system, a government registry — to look things up and to do things. A connector is the secure adapter that lets it. Think of it as a labelled, lockable socket: the workbench can only use the sockets you give it, only in the ways you allow, and every use is written to the record.

We build these connectors for your stack. One might let the workbench search your contract archive; another might let it file a regulatory return, post an update to your project tool, pull a counterparty from a public registry, or book a meeting and send the calendar invite — which is exactly what the assistant on our home page does when you ask it to set up a call.

A connector is the difference between an assistant that drafts the email and one that can actually send it — with your permission, and on the record.

Because they follow the MCP standard, each connector is a clean, swappable part: new ones can be added later without rebuilding the workbench, and you decide what every connector may touch. Nothing acts without the permissions and the human sign-off from the four commitments above.

§ 06

How working together actually works.

01 · 2–6 weeks

We listen

We sit with your team, find the one thing that would help most, and write a clear plan with a fixed price. $25–75K.

02 · 4–8 weeks

We teach it your business

We connect the sources, systems and workflows that show how your organisation actually works.

03 · 8–16 weeks

We build it

We build and set up your AI, with the safety checks and access rules your business needs. $150–500K.

04 · 1–2 weeks

Handover

We train your team and hand it over. You own it.

05 · ongoing

We keep it running

Updates, fixes and new features — or you run it yourself. $50–250K / year.

Most people leave this page without typing. You don’t have to.

Cleo, the design agent on the home page, sketches a private AI workbench for your business in a few minutes and emails the plan to you. It costs nothing, and commits you to nothing.

Talk to Cleo →