AI for Rail · LTA × Lorong AI · 21 July 2026

Scaling expert-level
maintenance with AI.

AI-powered inspection & diagnostics for the Lights-Out Depot. Today, you'll watch it work live.

CAPTURE DETECT DIAGNOSE DECIDE
SPEAKERVinoth Varatharajan · CTO, Singapore
TRACKLights-Out Depot · 1.1
CLIENTSGardens by the Bay · Mandai · NTUC Income · Julius Baer
Who we are

Monstarlab. 20+ years of
enterprise software, worldwide.

Tokyo-listed
~1,500 people
Global delivery centres
Singapore on the ground
TRUSTED BY Gardens by the Bay· Mandai· NTUC Income· Julius Baer· UOB

Serious software for serious organisations — digital products, platforms and custom AI, delivered at enterprise scale for Singapore's leading names.

Monstarlab global presence: 29 offices, 1,400+ employees across Asia, Europe, Middle East and the Americas
The problem · LTA's challenge 1.1

Same wheel. Two inspectors.
Two different outcomes.

DAY SHIFT · VETERAN, 20 YRS Crack found → fixed NIGHT SHIFT · JUNIOR, 6 MOS ? Missed → fails in weeks
MAINTENANCE QUALITY DEPENDS ON WHO'S LOOKING

A 20-year veteran glances at a train wheel and catches a hairline crack on the flange. The junior on the night shift checks the same wheel — and passes it.

The depot can't clone the veteran. And when he retires, that judgement leaves with him.

"How might we scale expert-level maintenance through AI-powered inspection and diagnostics?"

— LTA · Lights-Out Depot · 1.1
Why it matters

One missed crack becomes
a stopped train.

Missed defect inconsistent inspection Fault grows weeks in service Unplanned failure train pulled out Service disruption commuters affected
COST

Unplanned downtime

Reactive repair costs far more than a defect caught at inspection.

CONSISTENCY

Shift-to-shift variance

Quality swings by person, fatigue and experience — round the clock.

EXPERTISE

Retiring knowledge

Decades of judgement walk out the door with every retirement.

Our solution
Powered by MonstarX · our AI product suite

An AI that inspects like your best
engineer — every time.

01 · CAPTURE Camera scans each part 02 · DETECT crack · 0.82 Custom model flags defects 03 · DIAGNOSE Hairline crack on wheel flange. Severity: HIGH Replace within 2 service cycles. VLM writes the assessment 04 · DECIDE ⚑ FLAGGED threshold crossed Work order #4471 raised automatically Job created in your CMMS

NIGHT SHIFT OR DAY SHIFT · JUNIOR OR SENIOR · THE RESULT IS IDENTICAL

Live demo · 90 seconds
Powered by MonstarX

Watch it work.

INPUT · IMAGE / WEBCAM brake & wheel unit 34 ms DETECTION crack · 0.82 pad wear · 0.78 boxes + confidence, live DIAGNOSIS Hairline crack, flange + 78% pad wear. Severity: HIGH Replace in 2 cycles. → work order #4471 raised plain-English verdict + action
Under the hood
Powered by MonstarX

Built to run on-site,
in real time.

Camera rig 4K / line-scan EDGE GPU · ON-SITE · NO CLOUD CV detector custom-trained boxes + confidence VLM severity + action plain English CMMS work order raised Technician confirms / corrects FEEDBACK LOOP · CORRECTIONS RETRAIN THE MODEL
MOAT

Custom-trained CV model

Fine-tuned on rail defect imagery — not a generic API call.

LATENCY

Edge inference

Milliseconds per frame, fully on-site. No network dependency.

LEARNING

Human-in-the-loop

Every technician correction makes the model sharper.

Proof · Julius Baer

We trained a model to recognise
luxury goods — and price them.

For Julius Baer, we built a custom model that identifies luxury brands across watches, bags, cars and wines — then analyses the likely price range from the image.

See the object

Recognise brand + product from a photo.

Reason about it

Estimate its price range — brand there, severity here.

Deployed in production

The same muscle rail inspection needs.

Watch Bag Car Wine Custom model trained by Monstarlab Brand: ✓ identified Price range analysed
SEE IT → UNDERSTAND IT → ACT ON IT · SAME MUSCLE AS RAIL
Proof · Retailetics · Malaysia
Built by our Singapore team

We built the machine that
teaches the model to see.

DCM data collection machine with eight cameras, turntable and touchscreen, ready to capture a product

For Retailetics, we built the DCM — Data Collection Machine. When a new product arrives at the warehouse, staff place it inside.

Multi-camera 360° capture

Cameras + turntable image the product from every angle.

Barcode scan ties the label

Every image set is linked to ground-truth product data.

Sent straight to training

Each capture becomes labelled data for the recognition model.

SAME PRINCIPLE AS RAIL · MULTI-ANGLE CAPTURE → TRAINING DATA

Proof · Retailetics · Malaysia

The same model, live in
stores today.

The Smart Trolley carries a camera, barcode scanner and weight scale, running our own computer-vision model — deployed in stores today.

Scan & drop

Shopper scans the product and drops it in the trolley.

Vision match + confidence score

The model recognises the product and scores its confidence.

Weight cross-check validates it

Scale weight confirms the vision match — same logic as our rail "Decide" step.

Smart trolley deployed in a supermarket aisle, screen showing myCart by retailetics, scan here instructions
Smart trolley hardware design render with camera, screen and bin

DEPLOYED IN STORES · MALAYSIA · VISION + WEIGHT = VALIDATED CONFIDENCE

Trusted by

Real clients,
real deployments.

Monstarlab client logo wall including Google, Disney, UOB, Gardens by the Bay, Hitachi, Mandai, Julius Baer, Santander and more

ENTERPRISE DELIVERY ACROSS FINANCE · RETAIL · HOSPITALITY · PUBLIC SECTOR

Our AI product suite · MonstarX

We ship tooling,
not just models.

MonstarX

Our AI studio — a team of agents that builds complete applications.

MonPage

Agentic page & app generation, from prompt to product.

MonGPT

Our LLM layer powering assistants and reasoning across products.

CUSTOM MODELS + AGENTIC WORKFLOW + ENTERPRISE DELIVERY · UNDER ONE ROOF

The path for LTA

Your data in → a Singapore-
MRT-tuned inspector out.

TIER 1 · TODAY Base model pre-trained · generic + rail data TIER 2 · HOURS Rail fine-tune public rail-defect datasets TIER 3 · THE PILOT MRT fine-tune a few hundred labelled depot images from one depot → measured accuracy → in ~2 weeks

TRANSFER LEARNING · SMALL DATA ASK · FAST, MEASURED RESULTS

Why Monstarlab

What a pure-LLM team
can't bring.

MOAT

Custom-trained models

Proven ability to train domain models — not just prompt an API.

HARDWARE

Edge + sensor capability

DCM & Smart Trolley prove we build the physical side too.

TOOLING

MonstarX product suite

Agentic tooling that accelerates delivery end-to-end.

SCALE

Enterprise backbone

Tokyo-listed, ~1,500 people, 20+ years of delivery.

PROVEN

Production systems

Julius Baer, Retailetics, NTUC Income, Mandai and more.

LOCAL

Singapore team

On the ground, ready to work alongside LTA and operators.

Beyond 1.1

One engine — the whole
living railway.

1.1 · DEPOT — TODAY Inspection & diagnostics expert-level maintenance at scale 2.2 · MAINLINE Real-time defect detection same engine on track footage 3.3 · STATIONS Safety & access detection from existing cameras

START AT THE DEPOT · GROW ACROSS THE NETWORK

The ask

Let's run
a pilot.

Give us a small labelled image set from one depot. We'll return a Singapore-MRT-tuned inspector with measured accuracy — in about two weeks.

Start the pilot
vinoth.varatharajan@monstar-lab.com
· AI FOR RAIL
EN · English
01 / 16
01 / 16
← → Navigate  F Fullscreen