Research,in production.

Selected engagements across medical imaging, agriculture, infrastructure and financial services, each setting out the challenge, the system we delivered and the measurable outcome.

Infrasel

Infrasel maintains roads in Malaysia. Together we built ROAM, an AI system that reads every patrol video, drone flight, complaint and claim, and flags what needs action.

70 TBDashcam and drone footage analysed
+14%Precision from clustering-based look-alike filtering
DaysTo cover a full road network that once took months
~8,000km² of Selangor in scope

The challenge

  • Slow, unscalable inspections

    Reviewing footage by hand across wide districts meant small defects went unnoticed until they were costly.

  • False alarms, fragmented data

    Manholes, patches and shadows were mistaken for defects, and records were spread across spreadsheets and chat apps.

  • Undetected drainage failures

    Blockages that cause flooding were hard to find in large, hard-to-reach areas.

What we built

ROAM · built with Infrasel

A centralised road management and operation system

One system across the whole job: find the damage, fix it on time, and report it with proof. Ten features in three phases, linked with the systems the teams already use.

[01/05] FIND // ROUTINE

From reactive to proactive

Potholes, cracks and faded road lines are found on patrol, before anyone complains.

  • Defects flagged automatically

    AI checks every patrol video. No one has to watch the footage.

  • Measured, no tape measure

    Width, depth, GPS position, road name and KM for every defect.

  • Know which roads to fix first

    Roads ranked by number of defects, so crews go where the work is.

  • Less time on site

    Officers confirm each find on screen instead of driving out to look.

Damage is found and measured before anyone complains.

Screens are illustrative: names, roads and numbers are not real client data.

Koning Healthcare

Koning makes FDA-approved 3D breast CT scanner. We partnered with them to add AI-driven image enhancement and automated analysis to their imaging pipeline, so clinicians no longer have to choose between a safer scan and a clearer one.

3×Clearer images at 300 projections, matching 1,500-projection quality without more radiation
89%Faster scan registration, with better alignment across patients
74%Faster reconstruction, for higher throughput and lower cost per scan

The challenge

  • Safety versus quality

    Lower radiation doses reduce patient risk but produce noisy images that can hide important detail.

  • Manual, inconsistent diagnostics

    Radiologists reviewed calcifications by hand: repetitive, fatiguing and variable.

  • Limited data, misaligned scans

    Privacy rules limit annotated datasets, and positioning differences make scans hard to compare over time.

What we built

Low-dose input vs AI-denoised
Raw · 300 projectionsAI-denoised

3× clearer at 300 projections · move across to compare

Detection pass

Objective size and count on every scan

Synthetic acquisition

Training data without new patient scans

Scan-to-reference alignment Misaligned · 6.2 mmAligned · 0.3 mm

89% faster registration

Reconstruction workload −74% reconstruction time

Full use of GPUs and multi-core CPUs

Bitwise Agronomy

Bitwise uses AI analytics to manage orchards. Its GreenView platform measures fruit yield and ripeness so growers can plan harvest, labour and storage. We rebuilt the vision models and the infrastructure underneath them to run at orchard scale.

3.7B+Fruit detections, 2022–2024
19,477Kilometres of orchard footage processed
40%Faster inference
>70%Forecast accuracy, where traditional methods can miss by more than half

The challenge

  • Unreliable forecasts

    Traditional yield estimates can be off by more than 50%, disrupting harvest and storage planning.

  • Manual counting doesn't scale

    Hand counts are slow and miss the variation across very large orchards.

  • Difficult footage, heavy days

    Shifting light, dense leaves and camera movement, with growers uploading 500 GB+ on peak days.

  • Scattered insights

    Counts, ripeness, maps and system health lived in separate tools.

What we built

Upload volume vs workers

Scales out on heavy days, back in when demand drops

Orchard row scan

3.7B+ fruit detections, 2022–2024

Orchard block map

One monitored view · low-yield zones flagged

Model release pipeline

Zero-downtime model switch

MPay

MPay is a Malaysian eKYC provider. Its Trustgate verification depends on reading identity cards that users photograph themselves: tilted, blurred, glared or half-covered by a thumb. We built the module that straightens every card before it reaches OCR.

100%Homography success on the held-out test set
1MReal-world images the model was tested and validated on

The challenge

  • Poor image quality

    Tilted, blurred, cropped or cluttered photos lowered OCR accuracy and added manual checks.

  • Failure in hard conditions

    Fingers on corners, glare and low contrast broke traditional contour methods.

  • Inconsistent orientation

    Rotated or flipped cards made OCR miss text regions and raised rejection rates.

What we built

Corner heatmaps 4/4 corners · confidence 0.98

All four corners, even under a thumb

Template match Best match · upright, unflipped

8 orientations scored

Homography
H = [ 1.02 −0.14 18.6 · 0.09 0.97 −7.2 · 0.0004 0.0002 1 ]

Every card warped to a top-down view

Glare recovery Keypoints · low confidenceSegmentation · recovered

Nothing fails silently

Field extraction

Six fields, structured for downstream checks

Identity match

Card, face and liveness verified

See it onyour data.

Get in touch and we'll walk through the approach closest to your challenge.

Contact us