SAP to open-stack ERP migration
Manufacturer, 500+ employees. Phased migration off SAP ERP onto an open stack with custom modules — run in stages while the plant kept working.
- 0 downtime
- 8 mo migration
- −40% on licences
Computer vision development for real cameras, not benchmark datasets
A model that scores 98% on a clean dataset can be useless on your cameras: different light, different angles, dirty lenses and a network that drops frames.
What we can show you
Computer vision
The dataset is the project
Model choice is a week. Data collection, labelling and evaluation on your own footage are the months that decide whether it works.
What a computer vision project actually consists of
Five parts, and the model is the smallest of them. Problem definition in business terms — what counts as a miss and what it costs. Data: collection from your cameras, labelling, and a held-out set nobody trains on. Model: usually an existing architecture fine-tuned, occasionally a cloud API. Inference: where it runs — edge device, on-premise server or cloud — which decides both cost and privacy. Monitoring: quality drifts when the season, the lighting or the camera changes.
We start with a proof of concept on your own footage, because accuracy figures from anyone else's dataset predict nothing about your warehouse, shop floor or gate.
When computer vision pays for itself
When a person is currently watching a screen and counting, when quality control is sampled and subjective, or when an error has a price you can calculate. If the process happens ten times a day, a person is cheaper — and we will say so before taking the project.
Signals a vision project is worth a proof of concept
- someone watches a monitor and counts or checks things;
- quality control is a sample, judged by eye;
- cameras are already installed and recording;
- a miss has a cost you can put a number on;
- there is archive footage or images to learn from.
Price
PoC $7,990, production from $18,900
Timeline
PoC in 2–3 weeks
Payment
50/50 by milestone
Comparison
Cloud vision API or a model of your own
The cheap answer is often the right one, and we test both on your data before recommending either.
Criterion
Cloud APIYour own model
Own modelCloud vision API
Time to first result
Days
2–3 weeks (PoC)
Accuracy on your specific conditions
Whatever the vendor trained on
Tuned to your cameras and lighting
Cost shape
Per image or per minute, forever
Build once, then compute you control
Privacy
Footage leaves your network
Runs on-premise or on the edge
Works without internet
No
Yes, on edge hardware
When it wins
Standard tasks, low volume, no privacy constraint
Specific conditions, high volume, footage that cannot leave
Process
How we get from footage to something you can rely on
Two numbers decide the project: measured quality on your data and cost per stream. Both are produced before production work starts.
Evaluation on your data, not a benchmark
A held-out set from your own cameras, an agreed definition of a miss, and a harness your team can re-run after any change. Published accuracy figures from vendors predict nothing about your conditions.
We do not quote an accuracy number before the proof of concept. Afterwards you get a measured one with the method attached — the same discipline as in our engineering report.
Cost per stream, modelled up front
GPU or CPU time per camera, per hour, at the frame rate the task actually needs. Most vision projects die on the second invoice, not on accuracy.
Edge inference changes that maths and we model both options during the PoC.
Data and labelling as a plan
Collection windows that cover night, weather and shift changes; a labelling workflow with agreed guidelines; and a review step, because inconsistent labels cap accuracy long before the model does.
Your data, your labels: everything stays in your accounts and is assigned to you at delivery.
Drift is monitored, not assumed away
Seasons change, lenses get dirty, someone moves a camera. Quality is monitored continuously with alerts, and retraining is a scheduled activity rather than an emergency.
Where footage cannot leave the building, inference and storage stay on-premise — see AI security audit.
Why us
Why vision work lands here
Four things we can prove, rather than four adjectives.
No accuracy promises before the PoC
Anyone quoting a percentage before seeing your footage is quoting someone else's dataset. We measure first and put the method in writing.
Cost per stream in the proposal
Compute per camera at the required frame rate, with the edge alternative priced beside it. That number kills more vision projects than accuracy does.
Your data, your model, your weights
Footage, labels, trained weights and code are yours, with IP assigned at delivery. Nothing about the model is held as leverage.
We will recommend the cloud API
If a standard cloud service solves it at your volume and privacy requirement, that is the recommendation — and it is a much smaller engagement for us.
Industries
Computer visionby industry
Vision tasks are domain-shaped: pallet and label checks in logistics, defect detection on a line, occupancy and safety in property and retail.
- IT companies and digital agencies
- Insurers and brokers
- Lenders and microfinance
- E-commerce
- Wholesale and distribution
- Manufacturing
- Freight and forwarding
- Clinics and medical centres
- Dental
- Car dealers and service centres
- EdTech
- Legal and consulting firms
- Equipment rental
- Real estate agencies
- Property developers
- Travel and events
- Beauty and wellness
- Construction and renovation
FAQ
Computer vision questions
Six answers that usually replace a discovery call.
Do we need our own model, or is a cloud API enough?
Often the API is enough and cheaper — if your volume, privacy rules and conditions allow. We test both on your footage during the proof of concept.
How much data do you need?
It depends on the task and the variation in your conditions; the PoC establishes it. Labelling effort is estimated in hours, not waved away.
Can it run without an internet connection?
Yes, on edge hardware. That changes the cost and hardware plan, so we model both options before you commit.
What accuracy can you promise?
None before the proof of concept. Afterwards, a measured number on your own held-out data with the evaluation method attached.
What about privacy and footage retention?
On-premise or edge inference where footage cannot leave, retention rules agreed in writing, and access control designed in rather than added later.
Who owns the model and the data?
You do: footage, labels, trained weights and code, with IP assigned at delivery and a documented chain of title.
2–4×
faster on the code-writing share
in 3 weeks
from idea to MVP
230+
projects delivered
17 years
of engineering experience
Work
Work we can show you
SMARTWILL
Digital inheritance service (Web3)
AIRDEVELOPPA
Indoor air quality finder driven by sensor data
CAR5
Car-sharing app
APP.DENTAL
Cloud CRM for dental clinics
NEXORA
Warehouse and logistics management platform
As-built docs
Digital as-built documentation
EX211
Telegram Mini App for currency exchange
SMARTWILL
Digital inheritance service (Web3)
Have cameras and a person watching them?
Send a sample of footage and what a miss costs you. You get a feasibility view, a cost-per-stream estimate and a fixed quote for a proof of concept on your own data.