Sunny Bytes Solution · Penang

20+ years on the factory floor. Now I build the AI for it.

I'm Cheong Kin Leong (Sunny). I find the tasks that keep going wrong, build AI tools that fix them in days, not months, and train your people to use them. On your own hardware, so your data stays in the building.

Tell me about your task Training for your team

An illustration: the numbers are examples. The real Beru is on YouTube.

20+ yearsElectronics manufacturing and test at Benchmark, Quantum and Certance
95%+ yieldon stable models, and zero catastrophic quality returns in 10 years
150–200 modelsand hundreds of testers led as test engineering lead
Hours, not monthsto build a vision check with AI. Mine took a couple of hours.
Client: a local engineering firm On-site in Penang & Kulim, or online English · 中文 · Bahasa Malaysia Runs on your hardware You keep the code
What you get in one person

I speak your floor's language, write the code, and keep it running.

01I learn your line fastNPI, FCT, ICT, AOI, yield, takt, RMA: I've lived it for 20+ years. Show me your process once and I understand the system, the terms, and where it hurts.
02I know the painEscapes, rework, late reports, re-typed forms, the one expert everyone waits for. I've stood at those stations and owned those numbers.
03I write the softwareI've built shop-floor tools since my Benchmark days: yield tracking, real-time dashboards, report generators, and the control software for an automated screwing station.
04I keep it runningTools don't stop at handover. I maintain, fix and extend them as your line changes. With AI, what used to take me weeks now takes hours.
Sunny Bytes Solution

I fix the tasks that keep going wrong.

20+ years in electronics manufacturing and test taught me where mistakes come from: tasks done hundreds of times a day, by tired people, where one slip is expensive. I find those tasks, build AI tools that fix them, and deploy them on your line or in your office.

01 · on the line

Vision checks

PainA missing screw or a wrong label escapes to the customer.

FixA camera and a PC at the station check every unit, beside the operator.

Beru · built in hours
02 · test

Test-program automation

PainWriting a flying-probe program by hand takes an engineer days per board.

FixGerber and pick-and-place files in, a test program out, checked by your engineer.

Meru · in use since Oct 2026
03 · the office

Paperwork that does itself

PainReceipts and MCs taped to A4, scanned, then typed in again by finance.

FixRead from a phone photo. AI reads, a person confirms.

Neru · tried by 20 staff
04 · the office

Inbox and document assistants

PainCustomer emails get missed; answers sit in PDFs or in one senior's head.

FixEmail triage, reply drafts, and answers from your own documents that cite the page. On a private model.

Eeru · Aeru
05 · quality

Line data you can read

PainYield data is scattered across AOI, ICT and FCT machines, and reports are made by hand.

FixOne yield view, plus a first root-cause draft for the engineer to check.

Overseer · prototype
06 · infrastructure

Your own AI server

PainCustomer drawings can't go to the cloud, and Malaysia's amended PDPA adds rules for sending personal data abroad.

FixA private AI server, sized, built and set up in your building. I ran a 100+ GPU farm for two years.

see the farm →
How I work
  1. WatchI stand at the station or the desk, count, time, and ask what goes wrong.day 1
  2. BuildA rough tool, fast, with AI writing most of the code.hours to days
  3. ProveIt runs beside your people until the numbers say it works.1–4 weeks
  4. Hand overIt runs on your hardware. Your team owns it, and I train them.then it's yours

Try the method yourself in three short, free chapters: find the task · build fast · count what changed

What you get at handover
  • The working tool, on your PC or server
  • The source code. It's yours.
  • A one-page guide in plain language
  • Training for the people who use it
  • A support period we agree up front
Your data
  • Happy to sign your NDA before I see anything
  • Runs inside your network by default
  • Nothing goes to the cloud unless you choose it
  • Least access: read-only first, a person approves anything that's sent
Is it worth fixing? Do the sum

Put in one of your tasks.

If the numbers don't work, I'll tell you not to build it.

—hours back a month
—RM saved a month

Assumes 22 working days, RM 20 per staff hour, and a tool that catches 90% of mistakes. Change every number →

Send me this task
Learn something in 60 seconds

True or false? Make your guess.

Six things about AI that surprise the managers and HR teams I teach. Guess first, then see the answer. One minute, no jargon.

01 · true or false?An AI model is basically a giant Excel sheet.
Mostly true

It's billions of numbers in grids. An 8-billion-number model would fill about half of the biggest sheet Excel allows (1,048,576 rows × 16,384 columns).Open one → chapter 9
02 · true or false?AI reads your text in pieces, not in whole words.
True

They're called tokens. A long or rare word can be two or three of them. Rule of thumb: 100 tokens ≈ 75 English words.See it → chapter 10
03 · true or false?AI knows the whole answer before it starts typing.
False

It writes one piece at a time, each time guessing what comes next. That's why it can talk itself into a wrong answer, and why a person should check.Let it guess → chapter 10
04 · true or false?AI learned to see before it learned to chat.
True

AlexNet won the ImageNet contest in 2012, ten years before ChatGPT. Factory vision is older than chat.The story → chapter 6
05 · true or false?You need a data centre to run useful AI.
False

Open models run on a single gaming GPU, or even a laptop. Your data never leaves the building. I run mine in my office.What one GPU does → chapter 15
06 · true or false?A custom vision check takes months to build.
False

Mine took a couple of hours, because AI writes most of the code now. The hard part is knowing which check matters. That takes 20+ years on the floor.Find yours → chapter 17
Guess all sixKeep going: 19 short chapters, free →
The _eru family · selected work

Real tools, used on real lines

Each tool gets a letter and an "-eru". These three are the ones factory people ask me about most.

Meru pad report: every part's pads found and numbered on the copper

Every part's pads found and numbered from pin 1, straight from the Gerber copper.

Meru run profile: where each second of a run went, before and after an upgrade

Every run is profiled. One upgrade cut a 112-part board from 573 s to 345 s.

Meru M · Gerber → flying probe
In use · betaPythonOpenCVOCRLocal LLM
Problem
Writing a flying-probe test program by hand takes an engineer days per board.
Built
A web tool that reads the PCB fab package (Gerber, drill, pick-and-place, IPC-356), finds every part's pads, builds the netlist and picks probe points. When it meets a new part, AI identifies it and an engineer approves it into a footprint library that keeps growing.
Result
A 112-part board runs through in under 6 minutes, then engineers review what's left. Running on the client's own GPU server and used by their engineers since October 2026, with 500+ automated tests behind it.

Beru on a production run: every unit tracked, every screw checked.

Beru B · vision inspection
Built in hoursOpenCVNo model training
Problem
A missing screw is easy for a person to miss on the hundredth unit, and a custom vision system usually takes weeks.
Built
A camera inspection station that tracks each unit on the line and checks it's built correctly. AI wrote it with me in a couple of hours, using classical OpenCV: no YOLO model to train and no dataset to label.
Why it matters
With AI, a working inspection tool for a new part is now an afternoon's work, not a project. That changes which checks are worth automating.
Neru reads a fuel receipt and fills in vendor, date, category and amount for a person to confirm

From the Lesson 3 demo: snap a receipt, Neru fills the claim, a person confirms.

Neru N · Telegram office assistant
In useTelegramOCRLocal LLM
Problem
HR notices lost in WhatsApp, receipts taped to A4, finance re-typing every claim.
Built
A Telegram bot that reads medical certificates and fuel receipts from photos, threads support-group chats and answers in English, Malay or Chinese, all on a local model. AI reads, a person decides.
Result
About 20 staff made 123 scored attempts in its first class session.
Aeru

A · Private document Q&A. Reads your PDFs on your own machine and cites the page for every answer.

Eeru

E · Private email triage on a local model: threads, urgency, customers, deadlines and draft replies.

Geru

G · Finds probe-able pads from a flying-probe camera photo alone, with no CAD or BOM. Prototype.

AI training

Training that ends with a working tool

Right now I teach "Let's Grow Together with AI", a six-month course for the staff of a local engineering firm: finance, HR, sales, service engineers and managers. Half are beginners. Classes are 20–30 people, two hours each.

A full classroom of around 30 students Explaining what is inside an AI model file A small-group session around a meeting table A live Beru missing-screw demo on the meeting-room TV
Every lesson is an appOne offline file with live games, chatbots and tools inside.
Open the black boxStudents open a real model file and watch a CPU race a GPU.
The class talks to the AIRequests go by Telegram to my GPUs and come back on the projector.
Real chores, not toysClaims, VLOOKUP, WhatsApp chaos. "Automate a chore, not a job."
Lessons become filmsEach lesson is also a narrated, code-drawn video for revision.
Courses
2 days · engineers

AI for quality, test & yield

For QA, test and process engineers: read test logs, draft 8D and failure-analysis reports, and build one small tool from your own data. Taught from the test floor.

1 day · office

GenAI for the factory office

For admin, planning, HR and QA clerks: a prompt library for five of your real documents, and safe habits with customer data.

8 half-days · teams

Spot, build, deploy

For supervisors and engineers: each team finds one repeat error and ships one tool on the line, with me beside them.

1 day · owners & IT

Private AI for SMEs

Data-law basics, a simple AI-use policy, and hands-on time with a model that runs on your own PC.

Half day · operators

AI basics on the phone

In Bahasa Malaysia or Mandarin: safe, simple AI for SOPs and shift handovers, on the phone they already carry.

3–6 months · whole company

Grow together with AI

Two hours a week, real chores from every department, and a tool shipped every few lessons. The course I run now.

Taught in English, Mandarin or Malay. On-site in Penang, Kulim and the north, or online. Need it HRD Corp claimable? Tell me early (see the FAQ).

Try the free 19-chapter lesson →
Why an engineer

Not just a trainer. Someone who's run the line.

A typical AI courseSunny Bytes Solution
In front of the classA trainer who knows the toolsAn engineer with 20+ years in electronics manufacturing and test
ExamplesGeneric office demosYour own forms, your own line, your own mistakes
FocusOffice productivity toolsTest, yield, quality, and the paperwork around them
WherePublic classes, often in KLOn-site in Penang, Kulim and the north
You leave withNotes and a certificateA working tool your team already uses
Your dataUsually cloud toolsCan run on your own hardware; nothing leaves the building
LanguageUsually EnglishEnglish, Mandarin or Malay
After the classIt endsI can build, deploy and support the tools
From my lab

If I can describe it, I can ship it.

Since 2024 I've used AI to build things that used to need a team. I also built and ran a 100+ GPU farm for two years, so I can size your AI server honestly.

Mission: Japan, a family-trip film made entirely in codeMission: JapanAn 8-minute film, drawn entirely in code A row of six GeForce RTX graphics cards from the GPU farmThe GPU farm100+ GPUs, built and run for two years Exploded 3D view of an AOI machineMachines in 3DService-training models made in Blender
Open the lab →
Portrait of Cheong Kin Leong
About

Engineer first. AI second.

I'm Cheong Kin Leong, Sunny to most people: a computer engineer with 20+ years in electronics manufacturing and test in Penang, and the founder of Sunny Bytes Solution. My ground is high-level assembly and functional test, from pilot to mass production to sustaining. I learned most of my programming, and all of the AI, on my own. I like simple working things over theory, and I keep my hands on the hardware.

2026 – now
AI consultant & lecturer (contract), a local engineering firm

The six-month AI course, Meru, Neru, Eeru and Beru.

2024 – now
Founder, Sunny Bytes Solution (Erulab)

Built and ran a 100+ GPU farm for two years. Now AI solutions, training and consulting for factories and offices.

2007 – 2024
Staff Test Engineer / Manufacturing & Test Lead, Benchmark Electronics

16 years taking products from NPI to mass production for customers worth up to USD 15M a year. 95%+ yield, zero catastrophic returns in 10 years, 60% less floor space on one redesigned line.

2004 – 2007
Test Engineer, Certance and Quantum

Tape-drive test, including moving the test division from the US to Penang.

Working together

Ways to start

Solution · 2 weeksRepeat-error hunt

I review one line and its office, cost the top repeat errors, and build a prototype for the worst one.

Ask about a hunt →
Solution · 3–6 weeksA line-side tool

One fixed-scope tool, such as a test-log analyser, a vision check or a defect-photo log, with training and a support period.

Tell me the task →
Solution · 4–8 weeksA private AI box

An assistant over your SOPs and specs that runs inside your building, plus a simple AI-use policy for your staff.

Ask about private AI →
TrainingHalf a day to six months

Hands-on AI classes for mixed teams, taught in English, Mandarin or Malay.

See the courses →
Part-time role2–3 days a week

Need an AI and smart-manufacturing engineer on site, without a full-time hire? Let's talk.

Let's talk →
Questions buyers ask me
Do you work on-site?

Yes. I'm based in Penang and work on-site across Penang and Kulim. Training and follow-ups can also be online.

What do you need from us to start?

One task, one person who does it every day, and a few hours at the station or desk. If a machine is involved, some sample files or images from it.

Does our data leave the building?

Not unless you want it to. I'm happy to sign your NDA first, and I build tools that run on your own PC or server.

Who owns what you build?

You do: the tool, the source code and the guide. If you'd like, I maintain and extend it as your process changes.

How long does a pilot take?

Usually two to four weeks from the first visit to a measured result. Small tools can be ready in days.

Which languages do you teach in?

English, Mandarin or Malay. My slides carry all three, so mixed teams can follow.

Can the training be HRD Corp claimable?

Tell me early. Claimable training must run through an HRD Corp-registered training provider with an accredited trainer, and the grant must be approved before the class starts. We'll plan the route for your company before we fix dates.

Are there grants for automation projects?

Sometimes. Malaysia has automation and digitalisation grants for manufacturers and SMEs, but the rules change often, so check the current terms. I can scope the project so it fits the grant you choose.

Can you join us part-time?

Yes. Two or three days a week as your in-house AI and smart-manufacturing engineer: building tools, training people, and keeping the tools running.

Sunny Bytes Solution · Penang · Kulim · Remote

Got a task that keeps going wrong?

Tell me about it in a few lines: what the task is, how often it's done, and what goes wrong. I'll tell you honestly whether AI can help, and what it would take.

erulab.my@gmail.com LinkedIn YouTube