Evidence, not slideware.

5M+ frames gastrointestinal video, self-supervised pretraining

• Self-Supervised Learning
• PyTorch
• Medical Imaging
• Vision Transformers

Context

An IRCCS hospital needed a foundation model for endoscopic polyp detection, trained on more than 5 million gastrointestinal video frames.

Constraints

Strict GDPR compliance and NDA boundaries, cloud-GPU cost ceilings, and a need for reproducible experiments across model architectures.

What I did

Designed the SSL pretraining strategy (masked image modelling with DINOv3-style self-distillation), built the cloud data ingestion pipeline, and set up experiment tracking and fine-tuning protocols for classification, segmentation, and detection.

Outcome

A reproducible foundation-model pipeline over 5M+ frames, with data governance designed with future EHDS requirements in mind and a path toward clinical validation and edge inference.

Real-time on Jetson on-device tracking and distance estimation, MQTT zone events

• NVIDIA Jetson
• Computer Vision
• PyTorch
• YOLO

Context

A ceiling-mounted system to monitor occupancy and inter-person distance, driving lighting cues when distance violations occur.

Constraints

Real-time inference on an NVIDIA Jetson, accurate distance estimation from a single top-view camera, minimal occlusions at 3-6 m mounting height.

What I did

Built a compact detection-and-tracking pipeline, calibrated the camera and applied homography to project detections onto the floor plane, and published zone events over MQTT.

Outcome

A working edge deployment performing real-time people tracking, distance estimation, and MQTT-driven lighting on constrained hardware.

Production deployment ROS2 / Nav2 coverage navigation on a commercial ride-on scrubber

• ROS2
• Nav2
• SLAM
• Lidar

Context

A commercial ride-on floor scrubber needed an autonomous indoor navigation stack for large facilities such as warehouses, retail spaces, and airports.

Constraints

Robustness to floor sheen, reflective obstacles and mixed lighting; human-safe behaviour; repeatable coverage routes; simple operator workflows.

What I did

Developed the ROS2/Nav2 navigation stack (SLAM, localization, and coverage path-planning) and validated it in simulation (CoppeliaSIM) and on the platform.

Outcome

Reliable autonomous navigation with repeatable coverage paths and smooth, human-safe motion in large indoor environments.

+9% grasp reliability industrial bin-picking, shipped in rc_visard and rc_cube

• Computer Vision
• Stereo Vision
• C++
• ROS2

Context

Roboception's stereo-based perception products power industrial pick-and-place. Grasping needed to be more reliable in real factory conditions.

Constraints

Challenging factory lighting and clutter, tight latency budgets, and deterministic production-ready behaviour for integration teams.

What I did

Designed and implemented core C++/Python software for rc_visard and rc_cube, improved grasp generation and validation in the rc_reason component, and tightened the sensing-to-planning data flow.

Outcome

Faster pick-and-place cycles and more stable grasp proposals under difficult lighting and clutter, improving grasp reliability by 9%.

12+ services on one k3s node health-gated, reversible Helm delivery with automatic rollback

• Kubernetes (k3s)
• Helm
• Caddy
• CI/CD
• PostgreSQL

Context

A personal Linux VPS hosting over a dozen services (analytics, workflow automation, CRM, dashboards, custom web apps), migrated onto a single-node Kubernetes (k3s) cluster, each packaged as a Helm chart.

Constraints

Zero-downtime, reversible cutovers; stateful databases with live data migration; automatic rollback; production incident recovery, all on one node.

What I did

Authored Helm charts (Deployments, Services, ConfigMaps, Secrets, PersistentVolumes, init-containers, readiness/liveness/startup probes), built health-gated delivery with `helm upgrade --atomic` and digest-pinned images, and ran live credential rotation and incident debugging (OOM self-healing).

Outcome

A reversible, self-healing platform with Caddy as the retained TLS edge, Cloudflare zero-trust access, and backup-first attended upgrades for stateful data.

Multi-machine ROS2 user study published in IEEE: bi-manual TIAGo with synchronised gaze

• HRI
• MoveIt
• ROS2
• Motion Planning

Context

A TU Wien research study on whether a robot's non-verbal behaviour makes its intentions legible to people during a collaborative task.

Constraints

Safe, collision-free motion around people; precise temporal coordination of head and arm; a controlled within-subject experimental design.

What I did

Built the interaction engine on the bi-manual TIAGo robot with MoveIt motion planning, synchronized gaze control, and ROS state machines, then ran the user study.

Outcome

Coordinated gaze-and-arm behaviour measurably improved perceived intention clarity, engagement, and interaction fluency (published, IEEE).