Custom AI
we sell the part you can measure. Not the part you cannot.
None of our AI work has ever run in a customer's product. We are saying that here rather than waiting for you to ask it after a deposit. What we sell is one narrow job that runs, handed over with its numbers — including when those numbers are “it did not work well”.
First: if you want to BUY a chatbot, this is the wrong page
A number of search terms lead people here, and most of them do not need what we do. If you want a sales chatbot, a document-scanning tool, or an AI product you can switch on this week — go and buy one. That is SaaS vendors' work, they do it far better than we will, and they charge a monthly price a business like yours can carry.
Names like MISA, Nhanh.vn, Base.vn or Ahachat in Vietnam, and Chatfuel internationally, are the right answer for that need. They have a product, they have a price, they are live in an afternoon. We do not compete there and we are not going to stand in front of you claiming we would do it better.
This page is for a different question entirely: “we have one specific process, our own data, and nothing off the shelf fits — is there a way, and how accurate is it actually?”
Where our AI work actually is
Three lines, with no rounding:
Yes
An MSc in Computer Science with an AI focus — machine learning, natural language processing and computer vision — from Can Tho University. Eight years of software development before that, mostly TypeScript and systems that actually run.
Not yet
None of our AI work runs in a customer's product today. The machine-learning side is applied research in progress, and we are moving it from coursework toward real work. We will not sell you a list of AI customers, because that list does not exist yet.
So what we are selling
Integration, measurement and operation. Putting a model into your actual process, measuring how right it is on your actual data, measuring what each run costs, then handing it over so you operate it. Not a pre-trained model, not fine-tuning a foundation model, not AI strategy with no build attached.
Why selling the measurement is the better sale
This is not a nice line to cover for missing experience. It is the technical reason: an AI feature with no numbers attached is an assertion, and an assertion you cannot check is not checkable until the day it runs wrong on your data.
Accuracy only means something on your data
An accuracy figure measured on somebody else's dataset does not transfer to yours — different industry, different way of writing, different noise. We build the evaluation set from your own samples, run it before handover, and report that number.
Cost per answer is not a constant
It depends on the real length of your sentences, how many calls you make a day, and what the model costs on the day it runs. We count on your own sentences rather than estimating from intuition — and we write the assumptions into the quote.
“It did not reach the bar” is a valid result
If the evaluation set comes out low, you should know that before you order the next package. A two-week measurement that answers the question is cheaper than a six-month build that reveals it was never going to work.
What we can honestly do
These two are the parts of our machine-learning work we have reason to believe in, and both rest on one thing: the Vietnamese text of a Vietnamese business.
Extracting Vietnamese documents
Contracts, invoices, handover minutes, stock notes — documents with a familiar shape scattered across half a dozen places. We do the part that reads them and lands them in a fixed schema, so the data can be used afterwards instead of re-read every time.
Normalising Vietnamese text into one schema
The same fact written four different ways is a defect waiting to happen, not a style preference. Getting it into one schema — and keeping it there after handover — is the actual work.
The entry package for AI work
Two weeks. It ends in a set of numbers on your own data — accuracy, cost per run, and what it gets wrong. Not a build.
The full AI and data serviceQuestions about AI
- Have you shipped AI for a client?
- Not in productised form. My machine-learning work is applied research in progress — an MSc with an ML/NLP/CV focus — and we are actively moving it from coursework into real work. We have no AI client list to put on this page and we will not invent one. That is exactly why the first package is a measurement rather than a build.
- Where should we start?
- With a measurement, not a project. The entry package is roughly two weeks and ends with numbers on your own data: accuracy, cost per run, and what the system gets wrong. You pay for the measurement before committing to anything further.
- What will it cost per month to run?
- we will not give you a number from the sky. It depends on your real sentence lengths, your daily call volume and the model's price on the day it runs — so we count on your own sentences and write the assumptions into the quote. Every AI quote we produce carries an estimated run cost line.
- Does our data get sent anywhere?
- No. We do not use your data to train somebody else's model, including our own. That is a boundary we write into the scope, and we think it belongs in writing.
- Do you do computer vision — product quality inspection from images?
- we do not sell it as a service. Image inspection needs a conveyor, controlled lighting and a real volume of defect images in the plant — things a plant with a line and daily quality checks already has. We describe that as a capability, not something you can order.
- Will this automate everything?
- No, and we do not promise it will. At a scale of a few dozen to a few hundred people, most of the time goes into data that disagrees with itself rather than into the model — a process where you cannot produce clean data will produce poor results, however good the model is.
Related pages
This page is not the whole offer. Most of the work we do arrives as an internal system first, with AI appearing somewhere inside it — or not appearing at all, in which case we say so.
Describe a process you do by hand and the outcome you need. We will say plainly whether AI is the right tool for it — and if it is not, we will say that too. The first conversation is online, carries no consulting fee, and does not require you to travel anywhere.
Send a process description