Full-stack AI engineering
We build AI into the daily work of Malaysian factories, warehouses, contractors and energy users. One team builds the data, models, agents, apps and devices it needs, then keeps them running.
Agentic AI
Agents read orders and RFQs from email, enter them into the ERP, draft the quote and chase late deliveries. Anything that spends money or reaches a customer waits for a person to approve it.
Physical AI
Cameras check each step before the unit moves on. Smart screwdrivers judge every screw by its whole torque curve. The models run on small computers beside the machine, so a slow network never stops the line.
Private AI
Language models that run on your own servers, in your own cloud account or in a Malaysian data centre. Your documents stay where they are, and every question and answer is logged.
Learning AI
Guided steps at the workstation, a camera that checks the work and tutors that explain in Malay or English. On one assembly line, new operators got up to speed 30–40% faster.
Most AI stops at the demo. Ours clocks in, on the production line, in the warehouse, on the tender and on the electricity bill.
AI works in your ERP, on the station screen and in WhatsApp. Nobody has to open a separate AI app.
It has to work at 3 a.m. with no engineer watching. We test it on your data, monitor it live and plan for what happens when it’s unsure.
AI drafts, checks and flags. People approve anything that spends money, reaches a customer or changes the line, and every decision is logged.
Before we build, we agree the numbers that count as success. Then we measure against them, before launch and after.
The full stack
AI projects stall on everything around the model: messy data, old systems, an app nobody wants to use, hardware on a noisy floor. We build all five layers ourselves, so nothing falls between vendors.
What people touch every shift: phones in the warehouse, screens at the station, dashboards in the office.
Agents and copilots that work inside your ERP, email and WhatsApp, with an approval step wherever it matters.
The right model for each job: a large language model to read a contract, a small vision model to check a screw.
Spreadsheets, ERP tables, sensor streams and PDFs, cleaned and joined so every answer starts from facts.
Small computers beside the machine for decisions in milliseconds. Cloud or your own servers for everything else.
What’s working now
Each one is running in production somewhere in the world today. We build them to fit Malaysian operations, and we start with the one that pays back first.
Agentic AI
Purchase orders, RFQs and invoices arrive as emails and PDFs. An agent reads them, checks them against the ERP and drafts the entry or the reply, and a person approves anything unusual. At Dow, an invoice agent caught a US$30,000 freight charge that should have been about US$5,000.
Agentic AI & Copilots
Vision inspection
A vision model checks every unit for missing parts, wrong labels and surface defects at line speed, and saves the photo with each decision. Audi uses the same approach to check about 5,000 spot welds on every car, instead of a sample.
Computer Vision & Inspection
Robotics
A camera tells the robot where each part is and which way it faces, so it can pick from a mixed bin without a precise jig. Robots that learn a new task from a few demonstrations are still in pilots, and we try them only where the payback is clear.
Robotics & Physical AI
Private AI
Open-weight models such as Llama, Qwen and Gemma run on a single GPU server and trail the best closed models by about four months. Contracts, drawings and HR files stay on your hardware or in-country, under your control.
Cloud, MLOps & Private AI
Learning AI
Digital work instructions show each step with photos and video, and a camera checks it was done right. In a study of 5,179 support agents, an AI assistant raised the output of new and less experienced staff by 34%.
Learning & Coaching AI
Research AI
Synthetic residents built from official statistics react to concepts, packs and prices before you pay for fieldwork. Studies show they rank options well but exaggerate the gaps, so we use them to cut ten ideas to two, then test those two on real customers.
Analytics & Research AI
AI-accelerated engineering
Language models read the codebase, find gaps in the test plan, draft automation scripts and group failure logs, and an engineer reviews every output. Airbnb moved 3,500 test files this way in six weeks, against an estimate of 18 months.
AI-Accelerated Engineering
The best way to predict the future is to invent it.
Industries
In Malaysia that means the factory, the warehouse, the tender office and the plant room. Pick an industry to see what AI does there today.
Guided assembly, camera inspection and smart tools that judge every screw.
Your RP4 bill rebuilt to the sen, and solar and batteries sized before you buy.
Labels read at the dock, stock that matches the shelf and proof of every delivery.
Every page of the tender read, and every panel counted and priced.
Launches rehearsed on a simulated Penang or Kuala Lumpur first.
Bilingual question banks, checked twice before a student sees them.
Try-on in the browser, stock you can trust and answers on WhatsApp at 2 a.m.
Every defect, work order and tenant request on record, and pumps that warn before they fail.
Capabilities
We don’t train foundation models. We build everything a model needs to do a real job: clean data, small task models, agents, apps and the servers they run on.
Results
Measured on our clients’ lines and in their offices after go-live. Client names stay private unless a client agreed to be named.
shorter visual inspection cycle with guided, camera-checked work
Guided work & inspection
fewer execution errors after paper instructions went digital
Digital work instructions
shorter quotation turnaround with AI drawing extraction
Drawing-to-quotation
saved every week on a digitalised production floor
DTC (M) Sdn. Bhd.
Results from individual BOE deployments. Yours will depend on your process and where you start. See all results
In production
Scenes from systems we’ve deployed. To protect our clients, some images are recreated, and names and products are removed.
What clients say
“We wanted to digitalize our production floor but were overwhelmed by the cost and complexity. BOE made it simple. They understood the process, customized everything we needed, and delivered under budget. It saves us at least 50 hours every week.”
“We did not expect much from a small company, but the final result blew us away. Everything was built exactly how we wanted. The team was fast, responsive and never tried to upsell us.”
How we work
The first consultation is free. Bring one real example, such as a tender, a job card, a year of electricity bills or a photo of the whiteboard.
We spend time with the people doing the job and write down every step, system and exception, including the ones nobody has documented.
You see it run on your own documents, photos or machine data, with a fixed proposal beside it.
We build it, connect it to your ERP, machines and apps, test it on real work and train the people who will use it.
We watch it in production, retrain it as your products change and add the next piece once the first one pays off.
Insights
Long reads on how AI is changing Malaysian work and life: what works, what failed and what it costs.
An indictment says a Malaysian official was toldA US indictment announced on 1 October 2026 says a co-conspirator emailed a Malaysian official in March 2024 that 27 Nvidia-equipped servers flown to Kuala Lumpur had been transhipped to a buyer in China. No Malaysian law then required a permit. Since July 2025 one does, and Malaysia has promised to share customs data with Washington.
When Johor’s taps ran dry, data centres trucked waterWhen silt cut water to 455,499 Johor accounts in late October 2025, a deputy minister said he understood data centres had bought water using their own tankers. Projects had asked for 808 million litres a day in 2024; the government now reports 63.9 million in use. Johor now requires reclaimed water wherever cooling needs it.
The AI-designed drug facing a 320-patient testIn September 2026 Insilico dosed the first patient in a Chinese Phase 3 trial of rentosertib, a drug whose target and molecule both came from AI, its developers say. The trial plans to enrol 320 people, and no regulator has approved the drug. Lilly’s deal with Insilico is worth “up to” about US$2.75 billion; the upfront part is US$115 million.
Talk to an engineer
Send one real example: a tender, a job card, a year of bills or a photo of the whiteboard. An engineer will map it with you and show you what AI can do with it. The first consultation is free.