01
Analyse the operation
We sit with the team and time the real work — the copy-paste, the chasing, the checking. The workflow that costs the most hours goes first, not the one that demos best.
03 — AI Transformation
Transform your operation.|Create impact.
BEAT 01
Every business runs on a handful of workflows that quietly cost it hours — a screenshot pasted into a chat, nine websites checked by hand, an order re-typed at a pickup counter.
BEAT 02
We analyse those workflows, build the right stack around them, and host it somewhere durable. Not a pilot — the thing your team uses on Monday.
BEAT 03
Three businesses below. Three operations that used to run on manual effort and now run on their own.
Who builds this
The practice lead.
Watcharatorn Chayapongplob
Kris · Founder & CEO, Yushi Marketing Technology
Ex-Google and ex-Amazon, now working at the point where AI stops being a talking point and starts changing how a business actually runs. The three builds below were scoped and led out of this practice.
- Speaker at GEN AI Masterclass, created by Digital IQ Academy with Marketing Oops
- Speaker at Explore AI Momentum, alongside 15+ AI specialists, and at Think with Google
- Hosts Web Wednesday TH — the 24.0 edition drew over 1,000 attendees
- Teaches AI for Business at Sripatum University, Bangkok University, The Connext and ThinkBerry Phuket
Completed the Agentic AI programme at MIT Professional Education, 2026.
Model agnostic
We are not married
We are not married
to one model.
Extraction, reasoning, classification and drafting are different jobs with different price and latency curves. We benchmark them per workflow and route each one to whatever wins — and re-route when the ranking changes, which it does.
Claude
Anthropic
GPT
OpenAI
Gemini
Google
Llama
Meta
Mistral
Mistral AI
Grok
xAI
DeepSeek
DeepSeek
Qwen
Alibaba
Open-weight models run in your own environment when the data cannot leave it. Where a hosted frontier model is the right call, we say so — and we tell you what it costs per run.
How we work
Analyse. Build. Host.
We are not selling you a model. We work out which part of your operation is bleeding hours, choose the tech and AI stack that actually fits it, build it, and keep it running in an environment that does not fall over.
02
Build the right stack
Vision models where the input is an image, scrapers and schedulers where it is a website, a proper database where it is a record. The stack follows the problem — never the reverse.
03
Host it durably
An automation that dies quietly on a Tuesday is worse than no automation. We run it in a monitored environment and stay responsible for it staying up.
Selected work
Three operations,
Three operations,
rebuilt.
Each of these started as a workflow someone was doing by hand, every day, and accepting as the cost of doing business.
01 · Distribution & delivery · in-house brand
ShinSen
Our own delivery business — so we felt this one daily before we fixed it.
Before · by hand
Staff screengrab the delivery table from their system.
The image is posted into a LINE group.
The data now exists only as a picture in a chat thread.
Nothing can be totalled, searched, or compared to last week.
After · agent
An agent watches the group and captures each report as it arrives.
It extracts the table from the image into structured fields.
Delivery data is stored properly, with history.
Management reads it as an interactive dashboard, not a photo.
The shift
The daily report stopped being an image someone had to scroll back for, and became data the business can actually query.
02 · Logistics & freight
Eastern Air Logistics
Vessel checking across nine separate port websites, done manually by customer service.
Before · by hand
CS staff open nine different port websites, one at a time.
Each vessel status is read off and noted by hand.
The client is updated only when someone gets to it.
Management has no live view of where anything is.
After · agent + TMS
An agent checks all nine ports automatically, on schedule.
Changes are pushed straight to the client as a LINE notification.
A TMS holds the shipment record end to end.
Management sees status live on a real-time dashboard.
The shift
Nine manual lookups became zero, and the client hears about their vessel before they think to ask.
03 · Commerce platform · co-owned
Pig Me Up
A celebrity pre-order platform where a wrong order is a public problem.
Before · by hand
Pre-orders collected and reconciled manually across channels.
Quantities and variants drift between the order and the pick list.
At the event, pickup is matched by hand against a list.
Every mismatch happens in front of a fan, at a counter, in a queue.
After · platform
A purpose-built pre-order platform owns the order from checkout.
That record stays authoritative all the way to fulfilment.
On-event pickup is verified against it, not against a printout.
Built and co-owned by us — we carry the outcome, not just the invoice.
The shift
Zero ordering errors and zero pickup errors — with a queue that long, on the day, in front of the people who care most.
The stack
What we build
What we build
these on.
Boring, proven technology chosen because it is still maintainable in three years — not because it was interesting to us this quarter.
Laravel
Application backbone, auth, queues and admin
Node.js
Agents, schedulers and integration services
Database
A real schema with history, not a spreadsheet
LINE notification
Where the client already reads their messages
Email notification
For the records that need an audit trail
Dashboards
Interactive management views over live data
The pattern
The output goes|where the work is.
PATTERN 01
Two of the three systems above deliver into LINE, because in Thailand that is where the conversation already happens. Nobody had to learn a new tool.
PATTERN 02
The best interface is frequently no interface. If the check runs on schedule and only the exception is surfaced, there is nothing to train anyone on.
PATTERN 03
Dashboards are for the people deciding. Notifications are for the people doing. Confusing the two is how good systems go unused.
What we build on
Five layers,
Five layers,
bottom up.
We refuse to start at the top. An agent sitting on unreliable data is just a faster way to be wrong.
Model-agnostic
Vision extraction
LINE Messaging API
Scheduled agents
Managed hosting
Monitoring
05 · Delivery
LINE notifications, dashboards, the client's own screen — where the work already is
04 · Agents
Scheduled, scoped, tool-using, with a human checkpoint on anything irreversible
03 · Extraction
Vision and parsing that turn screenshots, portals and PDFs into structured records
02 · Data
A real schema with history, so yesterday can be compared to today
01 · Hosting
Monitored, backed up and maintained — the layer most automations quietly skip
0
Port websites checked automatically instead of by hand
0
Ordering and pickup errors on the co-owned platform
0
Operations rebuilt end to end and still running
0
Place the output lands — the app your team already opens
Our position
"A screenshot in a chat group is not a report. It is a report you have not built yet."YMT · AI Transformation practice
Built for
Start here
Show us the workflow nobody enjoys.
The one with the screenshot, the spreadsheet and the person who checks it every morning. That is usually the one worth rebuilding.