- Home
- AI development
- Computer vision
Computer vision
Computer vision that turns photos and scans into records in your system
Computer vision is software that reads images: a scanned form, a receipt, a meter, a delivery photo, a shelf. We build it to turn those images into fields in your system, with a confidence check and a person for the photos it cannot read.
- Prices published on the site
- First working version in two weeks
- Support month to month, cancel any time
Two approaches cover most business work with images
A vision-capable language model such as Claude or Gemini reads documents and photos and returns fields: the vendor, the date and the total on a receipt, or the number on a meter. A trained detector is a smaller model taught on your labeled photos. It counts items or finds defects that a general model misses, and it can run on a modest computer near the camera.
We choose the approach from your photos, test it on photos the model has not seen, and show you the misses before anything goes live. Results depend on consistent photos and light. Clean scans of printed forms are the easy case; blurry phone shots in a dim truck cab are not.
Set up once, then every photo takes the same path
Set up
Collect
100 to 300 real photos or scans from the actual work, the bad ones included.
Label
Mark what matters on each one: the total, the dial reading, the damaged corner.
Choose the approach
A vision-capable language model to read documents and photos, or a trained detector to count items or find defects.
Test
Run it on photos the model has not seen and score it against what a person would record.
Every photo
Photo or scan arrives
From a phone, a scanner, an email or a shared folder.
The model reads or detects
Fields are read, items are counted or defects are marked.
Confidence check
Rules such as line items adding up to the total, two reads agreeing, or the detector's own score.
A person checks doubtful ones
The photo sits next to the model's reading, with the reason it was unsure.
Saved to your system
The record lands in your database, accounting system or work order.
A person reviews the low-confidence share. We lower it only as the log shows the model is right on that kind of photo.
Six jobs we build with computer vision
Scanned forms and receipts
Reads a scanned intake form, a W-9 or a receipt photo into fields, and flags the ones it cannot read.
Meter and gauge readings
Reads the number from a photo of a utility meter, a pressure gauge or a tank level and compares it with the last reading.
Delivery and job-site photos
Reads the label or work order number in the photo, files it under the right job and flags a photo too dark or blurry to use. Proof-of-delivery photos fit trucking and courier work.
Shelf and inventory counts
Counts items in a photo of a shelf or a pallet with a trained detector and writes the count next to the number in your system.
Damage photos
Outlines scratches, dents or water stains in photos taken for a returned item or a rental unit, for a person to confirm.
Equipment nameplates
Reads the model and serial number from a photo of the plate on an HVAC unit or a pump and fills the service record.
Which image jobs are worth automating
Worth building when
- Hundreds of photos or scans a month are retyped into a system by hand
- The images come from a consistent source: the same scanner, the same app, a similar angle and light
- A miss costs a re-check and nothing worse
- You can send 100 to 300 real examples, including the messy ones
Skip it, or keep a person, when
- The data can be captured at the source: a barcode, a QR code or a fillable form is cheaper and more reliable than reading a photo
- Photos vary widely in light and angle and nobody controls how they are taken
- The result stands in for an inspection the law requires: the model can prepare photos and readings, and a qualified person still inspects and signs
Priced by how many kinds of image and how many systems
One workflow running end to end, or one internal tool.
Several systems made to work as one: integrations, roles, reports.
Reading one document type, such as scanned intake forms or receipts, into one system usually lands in one process. A trained detector with its own labeled photos, plus integrations with an inventory or dispatch system, usually lands in connected systems. We estimate the model usage per photo in the audit from your volume.
Care $300 to $500 a month. Active $750 to $1,500 a month. AI model usage, GPU time and hardware are paid to the provider, on accounts in your name.
Questions about computer vision
How accurate is reading photos and scans?
That depends on the photos, so we measure it on yours. We test on 100 to 300 of your real images, bad ones included, and the confidence check sends anything doubtful to a person.
Do we need to train our own model?
Often not. A vision-capable language model such as Claude or Gemini reads documents and photos without training. A trained detector is for counting items or finding defects that a general model misses, and it needs a few hundred of your photos labeled by someone who knows what counts. We test the general model first.
Can it replace an inspection we are required to do?
No. It does not replace an inspection that the law requires, and we do not build it to. It can prepare photos and readings for the inspector or check that a photo was taken and is readable, and the qualified person still inspects and signs.
What do we need to send you to start?
100 to 300 real photos or scans from your actual work, and for each one the answer a person would write down. A folder of images and a spreadsheet are enough. The free audit tells you which approach fits and which price band it lands in.
Where do the images go?
The system can run on your own server. If it calls an outside AI model to read images, we tell you which one and what it sees before we build. Accounts and code repositories are opened in your name.
What happens to photos the model cannot read?
They go to a person. A review list shows the photo next to the model's reading and the reason it was unsure, and the person corrects it. We add those corrections to the test set, so later versions are checked against them. LLM evaluation describes how the test set works.
Related services
- Multimodal AI Text, images and voice in one system
- Language processing (NLP) Sort, extract and summarize text
- Trucking and courier Dispatch line and proof of delivery
- Local AI setup AI on hardware you own, set up on-site
- LLM evaluation Test which model does your job best
- AI development Every AI service we build, on one page
Tell us what's slowing you down.
You get a free written review by email within two business days.