Applied AI
Computer Vision & Multimodal AI
Vision systems for inspection, counting, document capture, and video understanding, on the edge or in the cloud.
Cameras are the cheapest sensor most businesses already own. We turn them into data: defect detection on a production line, stock and shelf counting, safety-equipment compliance, ID and document capture, and video understanding that finds the ten seconds that mattered in eight hours of footage. Multimodal models handle the messy, varied cases that classical vision never generalised to, while purpose-trained detectors handle the high-volume, low-latency ones.
What's included
- Detection & segmentation
- Document & ID capture
- Video understanding
- Edge deployment
What you get out of it
- Defects and exceptions caught in real time rather than at the end of a shift
- Manual counting and visual checks replaced with a continuous, auditable record
- Models that run where the camera is, so video never has to leave the site
From footage to a number you can act on
A vision project succeeds or fails on data, not architecture. We start with your real images (the badly lit, off-angle, partially occluded ones), build a labelled set that reflects them, and pick the smallest model that clears your accuracy bar at your latency and hardware budget. Then we ship it where it belongs: on a device at the edge when bandwidth or privacy demands it, in the cloud when it doesn't.
stream: line-3-cam-01 · 1080p @ 30fpsmodel: yolo-det (int8) on jetson · 41ms/framedetect: seal_defect conf=0.93 · frame 18422verify: multimodal check -> confirmedaction: reject_arm.trigger() · snapshot savedshift summary: 12,480 units · 27 rejects (0.22%)
What we build with vision
Detection, counting & segmentation
Object detection and pixel-level segmentation for quality inspection, inventory counting, occupancy, and process monitoring, tuned to your cameras and your lighting, not a benchmark dataset.
Document & ID capture
Photographed documents (receipts, forms, plates, meters, identity documents) cropped, deskewed, read, and validated, with liveness and tamper checks where the use case needs them.
Video understanding
Long footage summarised, indexed, and made searchable in plain language, so finding the relevant moment takes seconds instead of an afternoon of scrubbing.
Visual search & matching
Image embeddings for near-duplicate detection, catalogue matching, and visual similarity search across large product or asset libraries.
Edge deployment
Quantised models running on-device (Jetson, ARM boards, phones), so inference is instant, works offline, and keeps footage on site.
Technologies we work with
- PyTorch
- YOLO
- OpenCV
- Segment Anything
- ONNX Runtime
- TensorRT
- NVIDIA Jetson
- Python
- FastAPI
Frequently asked questions
Less than it used to be. Pretrained backbones and multimodal models mean a few hundred well-chosen examples per class often gets a usable first model, which we then improve with the cases it gets wrong in production.
Yes. For factory floors, vehicles, and remote sites we deploy quantised models to edge hardware and sync only results, which is usually cheaper and always more private than streaming video.
Almost always. Standard IP cameras over RTSP are enough for most use cases; we will tell you up front on the rare occasion the optics genuinely need upgrading.
We design for it: on-device processing, face and plate blurring where identity isn't the point, short retention windows, and documented data flows for LGPD and GDPR.
Next step
Need computer vision & multimodal ai?
Tell us what you're trying to build. We'll give you an honest read on the approach, the effort, and what it costs to run.