AI and custom model development
Assistants, workflow automation and custom models, connected to your documents and systems, tested against your own questions before they go live.
What we build
On your website or app for customers, or inside your team's own tools. They answer from your documents, FAQs and product data, and hand over to a person when unsure.
The repetitive reading and writing in a process: incoming emails, forms and documents are sorted and processed, with a person approving where it matters.
When a general model is not enough: fine-tuning on your examples, small task-specific models, and the evaluation sets that prove they work.
Search, summaries, drafting and suggestions built into the software your customers already use, designed with the same care as the rest of the product.
When a custom model makes sense
Most projects do not need a custom model. We start with the simplest approach that works and move along only when your own tests show a reason to.
Your documents and a set of real questions to test against.
Quality depends on how well your sources cover the questions. Gaps are found in testing and filled, or handed to a person.
How a model gets made
We map what you have: documents, past conversations, product data, labelled examples. We agree what may be used, and clean and structure it.
The simplest version first: a general model with your data. It sets the bar every later change has to beat.
Only if the baseline falls short: prompt and retrieval tuning, fine-tuning, or a task-specific model.
Every version is run against an evaluation set built from your own questions, and reviewed with your team before anything goes live.
Connected to your website, app or tools, with access control, logging and a hand-over path to a person.
Real conversations are reviewed, new questions are added to the evaluation set, and the model is improved when your data or needs change.
Answer or hand over
Pick a customer question for a sample online shop. See which sources are found, the drafted answer, and the decision the checks make.
Orders are dispatched on the next working day and delivered across Hong Kong, including Kowloon. You will get a tracking link by email once it ships.
ANSWEREDSent to the customer. The answer is fully supported by the sources.
Guardrails
Your data is used only for your project, under the terms agreed in the proposal.
When the sources do not support an answer, the assistant passes the conversation to a person instead of guessing.
Before launch, every version is run against an evaluation set built from your own real questions.
Who can use what is controlled, and conversations and actions are logged so they can be reviewed.
Refunds, complaints, legal or medical topics and anything sensitive can be set to need a person's approval.
| Question | Expected | Result |
|---|---|---|
| Opening hours on public holidays | Answer from FAQ | PASS |
| Change delivery address after dispatch | Answer, link to support | PASS |
| Discount not applied at checkout | Person approves reply | REVIEW |
| Complaint about a staff member | Hand over to a person | HAND OVER |
| Question about a product you do not sell | Say so, do not invent | PASS |
We use it ourselves
KoLink Chat
In Cantonese or English, and it hands the conversation over to the business's team when it is unsure.
KoLink platform
The KoLink platform uses AI for support answers and for summaries of campaign and brand reports.
Model-agnostic
We pick the model for each task on quality, speed, cost and where your data may go, and build so it can be changed later. Your data, prompts and evaluation sets stay yours.
Start a projectMore services
All servicesServices
Websites, web apps, and iOS and Android apps, designed, built and supported by one team.
View serviceServices
UGC-style videos, made by creators or with AI, produced video and motion design, delivered in the formats each platform needs.
View serviceFAQ
Usually not at first. Most projects start with a general model connected to your data. We only fine-tune or build a task-specific model when tests on your own questions show it is worth it.
It hands the conversation over to a person on your team, with the context attached, instead of guessing. Topics that always need a person can be set in advance.
Only for your project, under the terms agreed in the proposal. We agree which data may be used, where it is stored and which model providers may process it.
Gemini, OpenAI models, Claude and open-weight models. We choose per task and build so the model can be changed later without starting over.
Yes. Assistants can answer in Cantonese, written Chinese and English. KoLink Chat, our own product, already answers in Cantonese or English.
We build an evaluation set from your own real questions with the expected outcome for each one, and run every version against it. Your team reviews the results before anything goes live.
Yes. Assistants and automations can connect to your website, app, inbox, documents and internal tools, with access control and logs.
Contact
Tell us what you want to grow. We reply within one business day.