Reading
Structured fields from invoices, delivery notes, identity documents, forms and technical drawings. A hundred suppliers send a hundred layouts; we build a pipeline that does not depend on templates.
Service detail
Computer Vision Systems Development
Pipelines that turn what a camera or scanner sees into a record you can query. The images stay on site; only the result leaves.
Done Dynamics builds computer vision systems: document extraction, production-line quality control, camera analytics and plate recognition. We combine classical computer vision with models that read images, and deploy on an edge device or a server inside the building. The result: the video twenty cameras produce never leaves the network — what leaves is the record made from it.
Four families of work
Reading, counting, judging, watching
We separate these four at the start of every project, because their costs are nothing alike.
Counting
How many items on the shelf, how many pallets came off the truck, how many parts per minute on the belt, how many millimetres long. In a controlled scene, classical methods answer in milliseconds.
Judging
Defective part, unacceptable weld, an object that should not be there. The hardest family — consistency is only possible once the decision boundary has been written down.
Watching
Zone violations, safety equipment checks, vehicle and plate tracking, occupancy measurement. Producing results instead of recordings keeps the retention burden low.
2 minutes · 4 questions
What would computer vision earn you on your floor?
Four questions to see whether there is a real use case. If there is not, we say so.
How we build it
Images everywhere, data nowhere
The cameras run but nobody watches; the archive is full but cannot be searched. The job is closing that gap, not adding one more screen.
-
Images stay on site
Deployment runs on an edge device or a server in the building. What leaves the network is not video but a result record of a few hundred bytes.
-
Measure first, deploy second
A one to two week accuracy study reports what accuracy this job can reach with this method. If it falls short, the project stops there.
-
Classical methods and models together
Classical computer vision is faster and cheaper in a controlled scene; a model wins in an uncontrolled one. Most pipelines end up with both layers.
-
The error balance is your call
Passing a defect is not the same as rejecting a good part. The threshold is a business decision, set together and left adjustable.
-
Drift is monitored
Lighting shifts with the season, a lens gathers dust, packaging changes. Periodic accuracy checks are part of maintenance.
-
Compliance built in
Where staff appear in frame, notice obligations, retention period and access rights are defined with the deployment.
Use cases
Pipelines we have built on site
From accounting documents to the production belt, from warehouse counting to the front gate.
Invoice and Delivery Note Reading
From scanned page to accounting record: field extraction, total and tax consistency checks, uncertain records handed to a person.
Production Line Inspection
Where defects are rare, the system learns what normal looks like rather than what failure looks like. Every decision is stored with the frame behind it.
Warehouse and Dispatch Counting
Pallet, carton and label counting; missing-item detection before dispatch and automatic writing into the warehouse system.
Plate and Vehicle Tracking
Plate recognition at the gate, entry and exit records and barrier integration — the record stays on site.
Workplace Safety Monitoring
Helmet, vest and restricted-zone violation detection. The goal is an alert at the moment of the event, not tracking individuals.
Archive Digitisation
Turning a scanned but unsearchable archive into something queryable, classified and labelled by document type.
Anatomy of a pipeline
The model alone is not enough
Most of the accuracy is won before and after the model: preparation, type detection and consistency validation.
- Image preparation: deskewing, noise removal and cropping the region of interest — skip it and the accuracy of every later step drops.
- Determining the document or object type; searching for every field everywhere is both slow and error-prone.
- Field extraction with a separate confidence value produced per field.
- Consistency validation: do the line items sum to the total, is the date in a plausible range, does the tax rate hold.
- Automatic routing of anything below the confidence threshold to a person.
- Writing the result into the existing system and linking it back to its source — the document or the frame.
- Periodic measurement against a held-out validation set, with drift tracked over time.
Our own measurements
The vision tower costs memory, not speed
We measured four vision-capable builds on the same machine. Generation throughput is almost identical to the text-only builds; the difference shows up in peak memory. What decides whether a machine can process images is memory rather than the processor — the first table we look at when planning hardware.
| Build | Generation | Peak memory |
|---|---|---|
| 4-bit | 38.9 tok/s | 19.2 GB |
| 6-bit | 29.2 tok/s | 27.0 GB |
| 8-bit | 23.1 tok/s | 34.7 GB |
| bf16 | 13.2 tok/s | 55.8 GB |
Vision-capable builds; Mac Studio M3 Ultra, 512 GB unified memory, a multimodal prompt. Single run — an order-of-magnitude guide, not a benchmark.
Process
Five stages from samples to production
Measurement comes before deployment; if the result falls short, the project stops there.
- 01
Data Collection
A sample from the real distribution: blurred, rotated and badly lit frames included. A study built on clean examples misleads in production.
- 02
Accuracy Study
A measurable result before any deployment. One to two weeks, ending in a report. If it falls short, those are the cheapest two weeks you will spend.
- 03
Site Deployment
Camera placement, lighting and processing hardware. The intervention that saves the most time is usually not software but a correctly placed lamp.
- 04
Integration
Where the result is written, who gets alerted, how long the record is kept. A system on its own screen goes unused.
- 05
Monitoring and Maintenance
Periodic accuracy checks, camera maintenance, tuning as new case types arrive, and reporting.
Commercial model
Accuracy report first, investment second
In vision projects, cameras, lighting and mounting are a real hardware investment. So we split the work in two: a fixed-price accuracy study first, then a deployment contract based on its result. The report states what accuracy this job can reach with this method, and the hardware decision follows that figure. Maintenance is written as a separate line, because without periodic accuracy checks a system degrades quietly over time.
Related services
Enterprise AI development →
Deployments that combine what the images produce with assistants and automation on the text side.
Custom software →
Desktop and field applications that talk directly to barcode readers, scales, cameras and serial ports.
High-volume data storage →
Tiered storage for camera recordings and image archives, with a retention policy to match.
Outsourced IT support →
Ongoing maintenance for edge devices, cameras and the on-site server.
Ready for your next software project?
Book a free 30-minute discovery call with our team.
