Home / Use Cases / Computer Vision Production
Use Case · Edge & Perception AI

Computer vision that works in production.

We take CV systems from proof of concept to production deployment: inference pipelines, edge optimization, annotation and retraining pipelines, and drift monitoring. Reference deployments: Hayden AI (smart-city perception at municipal scale) and Cargomatic (dock and container CV for logistics).

→ Why CV projects stall before production

Computer vision POCs succeed in controlled conditions and fail in the real world. The model performs on the test set. It degrades on live camera feeds with variable lighting. It works on the GPU-equipped lab machine and can't run on the edge device in the field. It hits 95% accuracy on the initial dataset and slowly drifts as the physical environment changes — and nobody knows until users start reporting incorrect outputs.

The production gap in CV is almost never the model architecture. It's the surrounding infrastructure: the inference pipeline that can sustain the required throughput, the edge optimization that makes the model run on the target hardware, the annotation pipeline that creates a feedback loop from production errors, and the drift monitoring that catches degradation before it causes failures.

How we take CV to production

1. Data and annotation pipeline

Training data audit and gap analysis — what's in the dataset vs what the model will see in production. Annotation pipeline design for the specific object classes and edge cases that matter for your use case. Active learning integration to prioritize annotation effort on high-value examples from the production stream.

2. Model development and optimization

Architecture selection for the latency/accuracy/hardware tradeoff your deployment requires. Training, validation, and evaluation against production-representative data. Model optimization for the target environment: quantization, pruning, or distillation for edge; batch inference optimization for cloud.

3. Inference pipeline

Production inference infrastructure designed for your throughput, latency, and hardware requirements. Streaming inference for real-time applications (Kafka-backed, sub-100ms). Batch inference for high-volume post-processing. Edge inference with OTA model update infrastructure.

4. Integration with downstream systems

Output routing to the systems that act on CV results — control signals, alert pipelines, evidence packaging, analytics databases. The integration layer determines whether CV outputs are actually useful in the operational context.

5. Drift monitoring and retraining

Confidence score and detection rate monitoring in production. Golden dataset evaluation on a scheduled cadence. Retraining pipelines that incorporate production corrections. The feedback loop that determines whether the CV system improves or degrades over time.

[ The production requirement ]

CV that survives the real world.

Edge optimization
Target hardware
Model runs on the edge device that exists in the field — not the GPU workstation it was trained on. Quantization and pruning are built into every edge deployment
Drift monitoring
Continuous
Degradation is caught by monitoring before users report incorrect outputs — confidence distributions and detection rates tracked against baseline in production
Retraining loop
Production-fed
Production errors flow back into the annotation pipeline — the model improves from what it gets wrong in the real world, not just on static test sets
[ FAQ ]

Production CV — common questions.

What's the gap between a CV proof of concept and a production deployment?
A CV POC works on clean, labeled test data in controlled conditions. A production system works on real-world data under variable lighting, occlusion, and hardware heterogeneity — and doesn't degrade silently when conditions change. The production gap is: inference infrastructure (latency and throughput), annotation and retraining pipelines (so the model improves from production errors), drift monitoring (so degradation is caught before failures), and integration with downstream systems (control signals, alert routing, evidence packaging).
How do you handle model drift in production CV deployments?
Drift is the most common failure mode in production CV that teams don't design for upfront. Causes: seasonal lighting changes, new object classes, hardware replacement, dataset shift from environmental changes. Our standard approach: production monitoring that tracks confidence score distributions and detection rate changes, golden dataset evaluation on a scheduled cadence, and a retraining pipeline that incorporates production corrections. Drift is detected before it reaches the failure threshold, not after users report incorrect outputs.
What does edge CV deployment require that cloud deployment doesn't?
Edge CV adds constraints: power budget (limits model size), intermittent connectivity (requires on-device inference for the critical path), hardware heterogeneity (different device generations with different GPU/NPU capabilities), thermal management, and OTA update infrastructure. The model optimization work — quantization, pruning, or distillation to fit the edge hardware — is typically 30–50% of total CV engineering effort on an edge deployment.

CV system that needs to move from POC to production?

Tell us about your use case, the hardware environment, and where the current system breaks down. We'll tell you what the path to production actually looks like.

Schedule a conversation → Edge & Perception AI