Vision AI Team Lead with 5+ years architecting end-to-end edge computer vision systems and applied LLM verification layers. Pick an industry below to run a production sandbox simulation.
Multi-plate ANPR with blacklist routing, and plant safety-zone compliance.
LLM housekeeping assessment and strict PPE compliance pipelines.
Road defect & asset inspection, LLM false-positive filtering, drone patrol.
Edge deployments for urban analytics and plant mobility.
Custom YOLOv8 ANPR deployed on city CCTV edge nodes. A single scan resolves every plate in the frame, cross-references each against the blacklist database, and hands a confirmed hit to the city-wide re-identification service to reconstruct the vehicle's route.
Detects all plates in the current CCTV frame.
Vision AI monitoring inside the Perodua Smart Mobility Plant. Geometry alone only says "a person is inside a polygon" — an LLM reasoning layer reads machine state, proximity and interlock telemetry to decide whether that is actually a reportable violation.
Housekeeping assessment and high-precision PPE compliance.
Deployed at Petronas Research to monitor lab housekeeping. Drag the divider to set how much of the frame is in a cluttered state — an LLM reasoning layer (Qwen/LLaMA) reads the detected objects and returns a 5S condition assessment instead of a raw bounding-box dump.
Object detection and segmentation models (YOLO, PyTorch) customized for HSE and PPE compliance monitoring to ensure site safety standards are strictly met.
Road asset inspection and remote aerial patrols.
Dashcam and drone road surveys deployed in Malaysia, Singapore and Japan. Instance segmentation maps pothole surface area to the pixel, while a second head classifies roadside asset condition — signage, guardrails and road markings.
Raw segmentation over wet or shadowed asphalt is noisy — tar patches, manhole covers and puddles all read like potholes. I added a VLM verification pass (Qwen/LLaMA) that re-reads each crop in context and suppresses the ones that aren't real defects before they ever reach the maintenance backlog.
Verification layer off — all 6 raw detections forwarded to the maintenance backlog.
Wide-area aerial security patrols utilizing PyTorch detection pipelines. Designed to identify unauthorized access or structural anomalies from high-altitude moving camera feeds.
Who is behind the pipelines.
I build computer vision systems that survive contact with the real world — noisy CCTV, moving dashcams, aerial feeds, and industrial floors where a false alarm costs a supervisor's trust. Over the last 5+ years I have taken projects from data collection and annotation strategy through model training, edge optimisation, and on-site deployment.
The thread across my recent work is the LLM verification layer: pure detection models plateau on precision, so I pair them with vision-language reasoning that re-reads a detection in context before it becomes an alert. That single change is what turned several of these pipelines from demos into systems operators actually keep switched on.
Leads a 3–6 person data annotation team and owns internal labeling tooling for image, video and segmentation workflows. Architects edge CV deployments end to end.
Smart-city ANPR, Perodua Smart Mobility Plant safety compliance, Petronas Research housekeeping monitoring, national road surveys and drone patrols.
Completed while working full-time as a team lead.