Plant and method for machining a semiproduct
A machining plant for wood semi-products combining a safety apparatus with a computer-vision supervision system that detects dangerous operator actions the safety apparatus did not prevent.
/// Research
Papers & patents
Independent research done outside working hours, and the industrial IP that came out of the day job. Both share a bias: find the case where the thing breaks, then report that.
Independent research · December 2025
The Illusion of Robustness: How Generator Imbalance Undermines AI-Image Detection
Detectors for AI-generated images are usually reported as a single aggregate accuracy. That number hides a bias: how the training set is balanced across the generators themselves.
Key findings
- Holding features, model, preprocessing and test set constant and varying only the generator distribution swings SDXL accuracy from 35.7% to 82.9% — a 132% change.
- Spectral features generalise unevenly: modern generators (ChatGPT, Gemini, SDXL) reach roughly 80% unseen accuracy under leave-one-generator-out cross-validation, while legacy ones (DALL·E 2, MidJourney) fall to roughly 32%.
- The effect is feature-agnostic: a ResNet50 baseline shows the same sensitivity, so it affects detectors in general rather than one architecture.
- It names a distinct bias category — generator-level domain imbalance — separate from compression, resolution and semantic bias.
Independent research · November 2025
StoryHash: Lightweight Multi-Modal Fusion for Robust Video Retrieval
Video retrieval has to survive cropping, compression and speed changes without losing semantic accuracy. StoryHash fuses visual, temporal and structural signals into a single compact vector.
Key findings
- Combines visual (CLIP), temporal (action) and structural (scene graph) embeddings into one 527-dimensional vector.
- Reaches 98.6% Recall@1 and 0.9748 average similarity over 89 DAVIS videos, 14 transformation types and 1,242 tests.
- Ablation shows a dichotomy: action features add +28% similarity but only +0.4% Recall@1 — they raise confidence far more than detection rate.
- Lightweight temporal features (15 dimensions, 2.8% overhead) capture most of that gain, which supports efficiency-focused designs over expensive temporal models.
- Identifies a CLIP vertical-flip bias (84.3% detection) and shows rotation is the only transformation where action features help detection.
patents
Inventor or co-inventor on three software patents covering CNC machine control, CAD/CAM automation and machine safety. One remains under NDA; details are given only where disclosure is permitted.
Collision detection method for a machine tool
Collision detection for a machine tool control unit, comparing orthogonal projections of the tool and of the machine hood on a shared projection plane.
Panel optimisation (third patent)
Machine-control method registered during work on panel processing.