Skip to content
CRADIOTECH

AI Vision & Analytics

Production computer vision, from edge to enterprise.

We train, deploy, and operate vision systems that watch live video and turn it into decisions. They run around the clock in hospital operating rooms, on factory lines, on traffic cameras, and inside our own sports-analytics products. Not prototypes. Systems with uptime.

50+
AI models in production
90 FPS
custom detection on a Jetson Orin Nano
concurrent RTSP streams, real time
24/7
live deployments in the field

Capabilities

Everything between the camera and the decision.

Detection, tracking & counting

Custom-trained detectors that hold up outside the lab.

  • Custom object detection (YOLO family) trained on your data
  • Multi-object tracking with re-identification across cameras
  • People counting, occupancy & zone analytics
  • Vehicle analytics: counting, speed estimation, heatmaps

Pose & sports analytics

The discipline behind our own sports products.

  • 2D/3D pose estimation and action recognition
  • Swing, stroke & movement biomechanics from plain video
  • Automated highlight generation from full match footage
  • Ball and equipment tracking from a single camera

VLM video intelligence

Describe the event in plain English and the system watches for it.

  • Prompt-driven live-video monitoring (8+ vision-language models)
  • Zero-shot, open-vocabulary detection & segmentation with no retraining
  • Hot-swappable prompts over a REST API, mid-stream
  • Semantic image search across large libraries

OCR & document intelligence

  • License-plate recognition on fixed and mobile cameras
  • Structured field extraction from document photos
  • On-screen text detection for media monitoring
  • VLM and OCR pipelines that output clean JSON

Industrial quality control

  • Production-line defect and contamination detection
  • Worker-safety and zone-compliance monitoring
  • Multi-camera identity without duplicate counts
  • Alerting with escalation workflows, not just dashboards

Edge deployment

Where most vision projects die, and where we specialize.

  • NVIDIA Jetson deployment with TensorRT and INT8 quantization
  • DeepStream multi-stream pipelines
  • Multi-model stacks tuned into single-device memory budgets
  • Fleet monitoring, OTA updates & 24/7 operations
Vision stack
YOLODeepStreamTensorRTOpenCVSAM2Grounding DINOVision-language modelsPyTorch
Edge & infrastructure
NVIDIA JetsonINT8 quantizationDockerROS2KafkaREST APIsOn-prem GPUCloud GPU

Proof

Shipped and running.

A selection of production systems. Client names withheld by agreement. Every number is from a delivered system.

Hospital robotics program

Operating-room activity monitoring

A 6-node distributed vision-and-audio system watching live operating rooms. The full multi-model AI stack fits into 7GB of unified memory on a single edge device and runs around the clock.

NVIDIA Jetson6-node ROS27GB footprint24/7 in live ORs
Cradiotech product

MatchFlow: automated badminton highlights

Full-match video in, highlight reel out. Rally detection isolates the roughly 25% of footage that is real gameplay and cuts it into highlight films plus social-ready shorts.

Pose-based rally detection~25% gameplay extractionAuto highlight cuts
Security deployment

Custom weapon detection on edge

A single-class detector trained on a 70,000-image custom dataset, quantized to INT8 and served through DeepStream at 90 FPS on a Jetson Orin Nano.

70K training images90 FPS on-deviceINT8 / TensorRT
Traffic monitoring project

Real-time road-accident detection

Temporal video understanding over live RTSP feeds. Three concurrent camera streams analyzed in real time with no added latency, flagging accidents as they happen.

3× RTSP liveTemporal video modelZero added latency
Beverage manufacturer

Production-line contamination QC

Catching hair, string, and foreign objects inside bottles on a moving line. Motion segmentation, tracking, and classification chained into one deterministic inspection pipeline.

Inline inspectionTrack + classifyFactory-floor hardened
GCC traffic project

Mobile license-plate recognition

ANPR from 2K/4K mobile cameras using a cascaded detection, plate, and character pipeline, engineered for regional plate formats.

DeepStream cascade2K/4K mobile videoMulti-format plates

How we work

Pilot first. Then scale.

01
Discover

A short discovery of your workflow, data, and goals. No obligation.

02
Scope

A concrete brief: approach, timeline, and a fixed price.

03
Pilot

One real use case, built and live in weeks. Prove it before you commit.

04
Scale & operate

Expand across the organization. We keep it running, with SLAs.

Questions

Good questions, straight answers.

How much data do we need to get started?

Often less than you think. Zero-shot and open-vocabulary models can pilot with no training data at all. Custom detectors typically start performing usefully from a few thousand well-labeled images, and we can bootstrap labeling with model-assisted annotation.

Edge device or cloud, which should we use?

It depends on latency, bandwidth, and privacy. Live video usually favors on-site edge devices, since there are no streaming costs and data stays local. Batch analytics favors cloud GPUs. We routinely fit multi-model stacks into small Jetson-class devices, and we will recommend the cheaper path when it is the right one.

Who owns the trained models and the data?

You do. Models trained on your data are your deliverables, along with training code and deployment configs. We keep nothing proprietary between you and your system.

What accuracy can we expect?

We quote accuracy after a scoped evaluation on your footage, not before. Anyone promising numbers up front has not seen your cameras. Pilots exist exactly to establish real metrics on real data before you commit.

Can you take over an existing vision system?

Yes. We audit the current pipeline, stabilize it, and either optimize in place or migrate it to a maintainable stack, then operate it under SLA if you want it off your plate.

Put a camera to work.

Tell us what you need to see. We will scope a pilot on your own footage and have it running in weeks.