Try AI · Free

Five things AI can actually do, running on real data.

Not videos. Not screenshots. Working software you drive yourself — upload your own spreadsheet, point it at your own website, ask it your own question, and watch every step of how it got the answer.

Create a free account I already have one

Free, and no card One verification email, then you are in Your uploads delete themselves after four hours

Ask Your Documents · a real answer

“What are the payment terms?”

searching the passages it indexed

Net 30 from invoice date, with 1.5% monthly interest on anything outstanding.

MSA-2024-011 · page 4

Every answer carries the passage it came from. Click it and the page opens.

In the library today

Five demos, each built around a different hard part

Everything below is the real pipeline, not a picture of one. They call Azure, they run SQL, they stream audio over WebRTC. That is also why they sit behind a sign-in — and why the diagrams are the actual mechanism rather than decoration.

Ask Your Documents

Your files are read, cut into overlapping passages, turned into vectors and indexed. Ask a question and only the nearest passages become the answer — with the page it came from attached, so you can open it and check.

Full RAG pipeline

ONCE, WHEN THE FILES ARRIVE

Your filesPDF, Word, scans

Up to 20, deleted after four hours

ReadAzure Document Intelligence

Keeps the page each line came from

Split1,600 characters, overlapping

They overlap, so a sentence is never cut in half

DUCKDB · HNSW INDEX

Each passage becomes 1,536 numbers. Near in that space means near in meaning.

EVERY TIME YOU ASK

“What are the payment terms?”

The question is turned into 1,536 numbers the same way

The nearest five passagesfound in milliseconds — and only those

Nothing else from your files is sent anywhere

The answerNet 30 from invoice date, with 1.5% monthly on late sums.

MSA-2024-011 · p. 4

Click the citation and it opens the page it came from

An answer you can check. Nothing is quoted that is not in a passage you can open.

Private by Design

Names, companies, amounts and account numbers are swapped for placeholders before anything leaves your server, and swapped back after the answer returns. The model reasons about the structure and is never told who.

AI without handing over your data

NOTHING IDENTIFYING CROSSES THIS LINE

YOUR SERVER

THE MODEL · SOMEBODY ELSE’S COMPUTER

1 · Your document, as it isClaim filed by Marcus Webb of Blue Ridge Industrial for $48,200.

Names, companies, amounts, dates, account numbers — found by pattern, on your own machine

2 · What actually leavesClaim filed by [PERSON_1] of [ORG_1] for [MONEY_1].

3 · The model reads the placeholdersIt can reason about who claimed what, without ever being told a name. The structure is intact; the identity is not there.

4 · The answer, still in placeholders[PERSON_1] filed for [MONEY_1].

5 · Put back, on your sideMarcus Webb filed for $48,200.

The mapping from placeholder back to name never leaves your server. There is nothing to leak at the other end.

Spreadsheet Intelligence

It measures what each column means instead of guessing, says plainly which ones it cannot work out, then writes the SQL in front of you. Every number on the page has a derivation you can follow back to a row.

Text-to-SQL · AI that shows its working

FIRST, IT AGREES WITH YOU WHAT THE COLUMNS MEAN

Your workbookThree sheets, 263,562 rows, read in about eight seconds

Three real tablesorders · products · people

Every column measured, not guessed

Sales_Amount · measure

Region · category

Order_Date · date

Col_7 · cannot tell — you decide

A column it cannot work out stays out of every answer, rather than being quietly guessed at

THEN IT ANSWERS, IN THE OPEN

“Total sales by region, biggest first”

Typed, or spoken into the box

THE SQL IT WROTE — SHOWN, NOT HIDDENSELECT Region,
SUM(Sales_Amount) AS total
FROM orders
WHERE status <> 'returned'
GROUP BY Region ORDER BY total DESC█ran in 42 ms · 4 rows

Read-only, one statement, twenty seconds — and every number has a derivation you can follow

BY REGION

Most tools that promise to analyse a spreadsheet guess what your columns mean and never say which guesses they made.

Voice Agent

Point it at any website. It reads the whole thing in seconds, then you talk to it out loud. Nothing is turned into text in between — which is exactly why you can interrupt it mid-sentence and it stops, the way a person does.

Realtime speech-to-speech

FIRST IT READS A REAL SITE — SO IT IS NOT GUESSING

any-website.com

Your own site, a product page, a knowledge base

Deep dive

12 pages
48 sections
24,000 words

Follows its links too, then stops at a sensible size

What it now knows

Four ways we engage

Fractional CTO

Operational Intelligence Assessment

Blocked by bot protection? It says which service blocked it, and offers a site it can read instead

THEN YOU JUST TALK

You

Speech → text

A model

Text → speech

You

How voice is usually done. Four hops, and the gap between them is the pause you can hear.

You, speaking

your voice, as audio

its voice, as audio

Azure OpenAI Realtime, over WebRTCone connection, grounded in what it read

Nothing is turned into text in between. That missing step is why you can talk over it.

Point it at a URL and have a conversation about it ninety seconds later. No script, no intents, no call flow to maintain.

Intelligent Document Processing

Documents arrive wherever they arrive — a mailbox, a shared folder, an SFTP server. They are collected, kept, and read by a model trained on your document type, which gives every field a confidence score so you know which one to glance at.

Vision AI · custom-trained model

THEY ARRIVE WHEREVER THEY ARRIVE — COLLECTED CONTINUOUSLY

A monitored mailbox

HubSpot

A shared folder

An SFTP server

Collector

One repositoryThe original is kept, always

Nothing is lost, and nothing has to be filed by hand first

THEN IT READS THEM — AND TELLS YOU WHAT IT WAS UNSURE OF

One invoiceA layout the model was trained on, not a template you maintain

EXTRACTED, WITH CONFIDENCE

Invoice no.

99%

Total due

96%

Due date

61%

The low one is the useful one: it tells you which single field to glance at, instead of re-checking all of them

Into your systemsERP, accounting, CRM — as structured data, not as typing

A model trained on your document, not a generic reader hoping the layout has not moved.

Why we ask for an email

A straight answer, because it is a fair question.

These cost money to run

Every question calls a real model and real Azure services. An account is what keeps the library open rather than scraped flat in a weekend.

Your uploads stay yours

Each account gets its own workspace. Nobody else can reach your files, and the whole thing deletes itself four hours after you stop.

One verification email

We send a link, you click it, you are in. We do not sell the address and we do not add you to a list.

The library is still growing

What should we build next?

If there is something you have been told AI can do and you want to see it working before you believe it, say so. Leave a comment on LinkedIn — it reaches a person, not a queue, and the demos that get asked for are the ones that get built.

Ask on LinkedIn →

Seeing it work is the fastest way to know whether it is for you.

Every demo here started as something a client needed. If one of them looks like your problem, the conversation afterwards is a short one.

Talk to us See what we have delivered