Computer vision for everyday business: OCR, inspection and safety

Computer vision is now practical for ordinary businesses, not just the tech giants. A camera or a scanner, a well-scoped model and a clear definition of success can read invoices, inspect parts, count people on a work site or check a retail shelf — no research department required. The hard parts are not the algorithms; they are data labelling, honest accuracy measurement and deciding where the model should actually run.

Where does computer vision pay off outside big tech?

The projects that earn their keep rarely look like the demos at developer conferences. They look like a person doing a repetitive visual task — reading, checking, counting — that a model can do faster, at all hours, without fatigue. Four patterns come up again and again.

Document OCR and data extraction

Optical character recognition has been around for decades, but modern models handle what old OCR could not: skewed phone photos of receipts, handwritten fields on intake forms, tables in scanned PDFs, stamps and signatures overlapping the text underneath. The practical goal is rarely "read the characters"; it is "get the invoice number, date, vendor and total into the accounting system without a human retyping them." That framing matters, because it means you measure success per field, not per character, and you route low-confidence extractions to a person instead of guessing.

Quality inspection on the line

A fixed camera over a conveyor or a workbench can flag scratches, missing components, bad welds, mislabelled packaging or fill levels outside tolerance. The model does not need to understand the product; it needs to tell "normal" from "not normal" consistently under your lighting, with your parts. Inspection is one of the friendliest starting points because the camera position is controlled, the background is stable and every flagged item can be verified by a person on the spot.

Counting and safety on work sites

Construction and industrial sites use vision for headcounts in restricted zones, hard hat and high-visibility vest detection, vehicle counts at gates and alerts when a person enters the swing radius of equipment. These systems work best as an extra set of eyes for a safety officer, not a replacement — a flagged clip that a person reviews in seconds is more defensible, and more useful, than an automated verdict nobody checks.

Retail shelf checks

Photos of shelves — from a fixed camera, a staff phone or a field rep's audit app — can be turned into answers: is the product present, is it in the agreed position, is the shelf tag correct, roughly what share of the shelf does each brand hold? For franchises and multi-location retailers, this turns an occasional manual audit into a routine, comparable measurement across every location.

What does an accuracy claim actually mean?

Every vendor will quote you an accuracy figure. Before it impresses you, ask three questions: accurate at what, measured how, and on whose data?

For detection tasks, the honest numbers are precision and recall. Precision = true positives ÷ (true positives + false positives): of everything the model flagged, how much was real? Recall = true positives ÷ (true positives + false negatives): of everything real, how much did the model catch? A safety system tuned for high recall will raise more false alarms; one tuned for high precision will miss more events. Neither is wrong — but you should be the one choosing the trade-off, because you know whether a missed event or a nuisance alert costs you more.

Base rates matter just as much. If only a tiny fraction of parts are defective, a model that labels everything "fine" scores a wonderful overall accuracy while catching nothing. And OCR "character accuracy" is a near-useless figure for business documents: a model can read almost every character correctly and still get the total wrong on a great many invoices, because errors cluster exactly where it hurts — digits, dates and unusual vendor names.

A model that scores well on someone else's benchmark has told you almost nothing about how it will behave on your loading dock in November.

The only accuracy claim worth acting on is one measured on a held-out sample of your own images — your lighting, your camera angles, your paperwork — with the precision and recall trade-off stated openly.

The data labelling reality nobody budgets for

Ask anyone who has shipped a vision system what took longest, and the answer is rarely the model. It is the labelling: collecting example images, drawing boxes around defects or hard hats, transcribing document fields, and doing all of it consistently enough that the labels mean something.

A few realities to plan around. First, edge cases dominate the work — the crumpled invoice, the shelf photographed at a steep angle, the defect that only appears under the afternoon sun through the west windows. Your first labelled dataset will miss most of them, which is why serious projects loop: deploy to a small scope, collect the failures, label them, retrain. Second, labelling needs written rules. Does a partly occluded hard hat count? Is a scuff a defect or normal wear? Two reasonable people will disagree unless the guideline decides for them, and every disagreement becomes noise the model learns. Third, pretrained and foundation models have genuinely shrunk the labelled-data requirement — a task that once needed a large labelled collection can often start from a much smaller one — but they have not removed the need for a carefully labelled evaluation set. You cannot skip labelling the exam, even when you can shortcut the studying.

Plan for labelling as an ongoing operating activity, not a one-time setup cost. Products change, packaging changes, cameras get bumped, seasons change the light. A model nobody re-evaluates degrades quietly.

Edge or cloud: where should inference run?

Once a model exists, it has to run somewhere: on a device near the camera (edge), on servers reached over the network (cloud), or a mix. This is an engineering decision with real consequences for cost, privacy and reliability — our notes on cloud infrastructure cover the hosting side in more depth.

As a rule of thumb: safety alerts and anything on a flaky rural connection lean edge; document processing and shelf-photo analysis, where a second of delay is irrelevant, sit happily in the cloud; busy multi-camera sites often end up hybrid, with a small on-device model deciding which clips deserve a closer look. Privacy law adds its own weight — footage of identifiable people is generally treated as personal information under Canadian privacy law, and keeping it on site is often the simplest way to shrink your obligations.

How do you run a pilot that tells the truth?

The failure mode we see most is a pilot designed to succeed rather than to inform. A truthful pilot has a narrow scope (one camera, one document type, one line), a baseline (how accurate and how fast is the current human process, honestly measured?), an agreed metric chosen before anyone sees results, and a human fallback so that low-confidence cases go to a person from day one. Run it long enough to meet the awkward conditions — night shifts, month-end paperwork surges, winter light. If the pilot beats the baseline on the agreed metric, scale one step at a time and keep measuring; the second camera will surprise you less than the first, but it will still surprise you.

How OlDevs helps

Computer vision is one of the core disciplines of our AI development practice at OlDevs, a full-stack studio in Vancouver working with clients since 2014. We handle the whole chain — scoping the task, building the labelling process, training and evaluating models on your data, and shipping the surrounding software so results land in the systems your team already uses. We work in the open: a working demo every week, one accountable team from first call to launch, and you own all code, models, designs and IP outright. You can read about the process on our how we work page.

If there is a repetitive visual task somewhere in your business — a stack of paperwork, a production line, a job site, a wall of shelf photos — tell us about it. Request a quote and we will reply within one business day with an honest read on whether vision is the right tool, and what a small, truthful pilot would look like.