WGTECH is a small engineering team in London. We take a job your office does by hand, like checking supplier quotes or approving expense claims, and build software that does most of it. When the software isn't sure, it stops and asks one of your people.
Tell us about a jobSee what we take on
Here is one of those jobs from start to finish. The data is made up.
Waiting for you
A quotation, sent as a PDF
Example Supplier Ltd
Quotation Q-2291, dated 2 October 2026
40 × control panel CP-12 at £86.00 each: £3,440.00
40 × mounting kit MK-3 at £29.50 each: £1,180.00
Delivery to site: £250.00
Total excluding VAT: £4,780.00
Lead time: 3 weeks from order. Payment: 30 days from invoice. Valid until 31 October 2026.
What the software read
One thing for you to decide. The three lines add up to £4,870.00, but the quote states £4,780.00.
No model runs on this page. It shows the kind of review screen we build: point at a value on the right, or tap it, to see where in the document it came from.
We do five kinds of work, and we agree the scope and the price before starting any of them.
Software that reads what arrives, whether that is quotes, invoices, orders or spreadsheets. It pulls out the figures, checks them against your rules and enters them in your systems. Whatever the rules can't settle goes to a person, on a review screen like the one above.
At the end you have it running on your own systems, with the results from your own sample documents.
Leave, expenses, purchases and onboarding, handled in one place. Each request is checked against your policy and sent to whoever has to approve it. If it sits too long, the approver gets a reminder.
Every decision is recorded in your HR or finance system, and your team can change the policy checks themselves.
An assistant that answers from your own documents and shows which page each answer came from. It can do simple things, such as moving a booking, and it hands over to a person when it doesn't know. You also get the test questions we used and how it answered each one.
Whole products in which the AI is only one part. One team writes the Spring Boot services, the Vue web app and the native iOS app in Swift. The code goes into your repository, with tests that cover the failure cases and notes on deploying it.
For when a general-purpose model is too slow, too expensive or not accurate enough for one task. We fine-tune with LoRA, compress the model and set up self-hosted serving with vLLM. We build an evaluation set and measure a baseline first, so the comparison at the end is like for like.
Our quotes cover our own work. Model usage, hosting and third-party services are costed separately and agreed before we start.
Not every job is. The ones that pay back are done many times a week. They arrive as documents, forms, spreadsheets or messages, and they follow rules that can be written down, even when there are a lot of rules.
One more test decides most cases: a person has to be able to check the result faster than they could have produced it. If your job doesn't pass, we'll say so.
| What arrives | What the software does | What you end up with |
|---|---|---|
| Supplier quotes and invoices | Reads the lines and terms, adds up the totals | Rows in your comparison sheet or accounts system |
| Expense claims and receipts | Reads the receipt, checks it against your policy limits | An approved claim in your finance system |
| Leave and purchase requests | Sends each one to the right approver, reminds them if it stalls | A recorded decision and an updated calendar or order |
| Booking and service requests | Checks your policy and calendar, drafts the reply | A moved booking and a confirmation |
| Supplier spreadsheets | Cleans up the names, applies your category and price rules | Draft product listings |
| Questions from staff | Searches your handbooks and procedures | An answer, with the page it came from |
| Sales enquiries from your website | Sorts them, fills in the details, drafts a first reply | A lead in your CRM, assigned to the right person |
| Weekly figures from several systems | Collects and summarises them | A draft report, ready for review |
These are examples of work we would take on. They are not past clients.
We settle which job it is, where the software has to run, and the tests that will show whether it works. You send us a sample of real inputs.
We run an existing model over your samples. That tells us what is easy, what is hard and what the first version should cover.
Each milestone is working software that you can try. We review it against the agreed tests, and that includes the cases it gets wrong.
You get the code, the test results, the configuration and notes on running it. Ownership and support are set out in the agreement before work starts.
Wei Li runs WGTECH and leads the scoping and the AI work on every project. He has a PhD in computer science from the University of Birmingham, and he is co-founder and technical lead of ZoyMed, a medical AI company that works on surgical video.
His research is about making large language models smaller and cheaper to run. He has first-author papers on it at ICML 2025, NeurIPS 2024 and Findings of ACL 2025, and our model tuning work comes out of it.
Wei works with a second AI engineer and with backend, web and iOS developers, so one team builds the model side, the services and the interface. The people who will be on your project are named in the proposal.
WGTECH was incorporated in June 2026, and we have no client work to show you yet. That is why the example at the top of this page uses made-up data, and why we test on your own samples before we build anything.
Say what the job is and who does it today, which systems it has to work with, and roughly what your timeline and budget are. A few sample inputs help more than anything else. Take out whatever is sensitive first.
We'll reply with questions, or with a proposal that has a fixed scope and price.
For anything else, write to hello@wgtech.co.uk.