We build the thing
that does the work.

The automation itself. It reads the documents, applies your rules, asks a person when it is unsure, writes the result into your systems, and keeps a record of every decision. Built by a software company, tested against your real cases, and handed over with the code.

This is for you if

Services

Everything we build, in the order we build it

A first process usually takes six to twelve weeks. The long part is never the clever bit. It is the exceptions, the integration and the testing.

01

Write down exactly what it must do

Every rule, including what happens when the numbers do not match, when a document is missing, when the supplier is new, and when two people disagree. Signed off before anyone writes code. Arguments here are cheap and arguments in month three are not.

02

Design where it sits in your systems

What it reads, what it writes, what it must never touch, and what it does when one of your systems is down for an hour. Written as a diagram your own IT team can review and object to before we start.

03

Build the document reading

Pull the fields out of the paperwork and attach a confidence score to each one. Anything the system is unsure about is sent to a person instead of guessed. That threshold is a setting, so you can make it more or less cautious later.

04

Build the rules as a table your team can read

Your decision rules go in a table, not buried inside code. When a tolerance changes from two percent to three, someone in your finance team changes a number. They do not raise a ticket and wait a fortnight.

05

Build the screen where a person decides

A queue showing the document, what the system read, why it is unsure, and the options. Approve, correct or reject in one click. Every correction is stored, so you can see what it keeps getting wrong.

06

Connect it to your systems

Through a proper integration where one exists. Where it does not, we will say so and price the alternative honestly, because software that clicks through screens breaks every time your supplier changes their layout.

07

Test it against your own history

We replay real cases from the past year and compare what the system decides against what actually happened. Where they differ we find out which one was right. This is the only test that means anything.

08

Build the record of every decision

What came in, what the system read, what it decided, why, who approved it and when. Kept in a form an auditor can be shown. If you cannot explain a decision six months later, you do not really have a system.

09

Run it alongside your team before it takes over

For a few weeks it processes everything but changes nothing. You compare its output against what your team did. Only when the two agree consistently does it start acting on its own.

10

Hand it over properly

Code in your repository, written instructions, a guide for what to do when it breaks at midnight, and training for whoever owns it. You should be able to change suppliers without changing systems.

Describe one process on a call and we will tell you what it would take to build it.

Our stack

The tools, standards and methods we use

You do not need to know any of these. They are listed so your IT team can see what would end up in your systems, and check that none of it is unusual.

What we write it inOrdinary, widely used languages, so any developer can pick this up after us
PythonFastAPINode.jsTypeScript.NETCelery and Redis queues
What keeps the steps in orderHandles the waiting, the retries and the steps that take days, so we are not rebuilding that ourselves
CamundaTemporalApache Airflown8nPower Automate
Reading the documentsChosen per document type, based on what actually scored best on your own paperwork
Azure AI Document IntelligenceGoogle Document AIAWS TextractLayoutLMConfidence thresholds
The AI models, and how we keep them in lineModels are told to answer in a fixed format and only from your documents, so they cannot invent an answer
Anthropic ClaudeOpenAIAzure OpenAIFixed JSON outputRetrieval augmented generationPrompt version control
Proving it still works after a changeThe same test set runs on every change, so an improvement in one place cannot quietly break another
Evaluation setsRegression testsLangfuseLangSmithFallback models
Connecting to your systemsProper integrations first. Screen automation only where a system genuinely has no way in
RESTSOAPGraphQLOAuth 2.0WebhooksSAP, Dynamics, Odoo and ERPNext connectorsUiPath
Where it runsYour cloud or ours, with a test copy that is separate from the live one
DockerKubernetesAzureAWSGoogle CloudGitHub ActionsTerraform
Keeping it secureThe standard checklists, including the newer one written specifically for AI systems
OWASP Top 10OWASP Top 10 for LLMsSecrets managementRole based accessPenetration testing

Example

One invoice, from arrival to payment

This is the same supplier invoice process, followed step by step. It is worth reading because almost every question about cost and risk lands on one of these ten steps.

Example
09:02

The invoice arrives by email

The system watches the accounts inbox, saves the attachment, and records where it came from. Nothing is deleted, so the original is always there to go back to.

09:02

The fields are read

Supplier, invoice number, date, currency, line items, tax and total. Each field gets a confidence score. Here the total scores 99% and the purchase order number scores 71%.

09:02

The supplier is matched to your records

The name on the invoice rarely matches your master list exactly. It is matched on tax number and bank details rather than name, because names are written differently every time.

09:03

It is checked against the purchase order and the delivery note

Three way matching. Ordered 500 units, received 480, invoiced for 500. The system finds the difference rather than a person finding it three weeks later.

09:03

It stops, because two things are unclear

The quantity does not match, and the purchase order number was only read with 71% confidence. Your rules say anything under 85% goes to a person, so it joins the review queue instead of being guessed at.

11:15

A person decides

Your clerk sees the invoice, the order, the delivery note and the two problems side by side. She corrects the purchase order number and confirms the short delivery was agreed by phone. Two clicks, and both are recorded against her name.

11:15

The correction is kept

Her correction is stored as a labelled example. When the same layout appears again next month, that field is read correctly. The system gets better because she did her job, not because someone retrained it.

11:16

It is written into your ERP

Posted with the correct cost centre and tax code. If the ERP rejects it, the automation does not retry blindly. It stops and tells someone.

11:16

The approval is routed by your own rules

Under your threshold it goes straight through. Over it, to the department head. The rule sits in the table your finance team can edit.

Any time

The whole thing can be explained later

Six months on, an auditor asks why this invoice was paid short. One search shows the original email, what was read, what did not match, who approved it, when and why.

An example, with invented figures, to show how the pieces fit together. Your process, rules and thresholds will be different.

FAQ

Questions we get asked before starting

How long does a first process take?

Six to twelve weeks for most. What moves it is the number of exceptions and whether your systems have a proper way in. The second process is always faster because the surrounding parts already exist.

Will it replace our staff?

On the processes we usually build, it takes the repetitive part and leaves the judgement. The people who did the typing end up handling the exceptions. We will not tell you nobody's job ever changes, but a system that removes the review step is one nobody trusts.

What happens when it gets something wrong?

It should stop rather than guess. Anything below your confidence threshold goes to a person. When something genuinely wrong gets through, the record shows what it read and why it decided that, so it can be fixed rather than argued about.

Do we have to move to the cloud?

No. It can run in your own environment. It costs a little more to set up and some document readers are only available as cloud services, in which case we will tell you what the alternatives are.

Who owns what you build?

You do. Code in your repository from the first week, not delivered at the end. No part of it is locked to us.

Can you work with the system we already bought?

Often, yes. Many clients have a platform that does part of this and stalls on the integration or the exceptions. We are happy to build only the missing part rather than sell you a replacement.

Next step

Tell us the process and where it goes wrong.

What comes in, what your team does with it, where it ends up, and which part causes the arguments. That is enough for us to say on the call whether this is a build, a smaller fix, or something you should not automate at all.

Ahmedabad, India. We work with teams in the US, UK, Europe, Singapore and the Gulf, and we are used to the time difference.