Computer vision demos are easy to love. Point a camera at something, and a box appears with a label and a confidence score. Getting from that moment to a feature thousands of people use every day is a different job. We learned this building HAI, which starts every wellness check-in from two photos taken on ordinary smartphones, and from identity checks in our NexID eKYC platform.
Here's what changes after the demo, and what we now plan for from day one.
1. The real enemy is the photo, not the model
In the lab, images are sharp, well lit and centred. In the field, they're taken in dim rooms, at odd angles, through smudged lenses, by people in a hurry. Most errors we see in production come from poor input, not from a weak model.
So the first model we ship is often a quality check, not the main one. It runs on the phone, in real time, and tells the user to move closer, hold still or find better light before the photo is taken. In HAI, this check means the analysis starts from usable images, and it saves everyone a retake later.
2. Guided capture is a UX problem
Good capture is design work as much as AI work. Clear framing guides, short instructions, gentle feedback and an automatic shutter when the image is good enough all lift accuracy more than another week of training. This is where a mobile team that has shipped many consumer apps makes the difference.
If users struggle to take the photo, no model can rescue the result.
3. On-device or in the cloud?
We usually split the work:
- On the device: fast, small models for quality checks, cropping and liveness, so feedback is instant and works on weak connections.
- In the cloud: larger models for the main analysis, where we can update them without an app release and keep a clear audit trail.
Sending fewer, better images to the cloud also lowers cost and reduces how much personal data you move around.
4. Your data decides your accuracy
Public datasets rarely match your users, devices or lighting. The models that hold up in production are trained and tested on images collected the same way the product will collect them, labelled by people who understand the domain. That's why we built HAI's dataset ourselves, with consent and practitioner annotation. We wrote more about this in Your proprietary data is the real AI moat.
5. Plan for monitoring from the start
Phones change, cameras improve and user behaviour shifts. A model that was accurate at launch can drift quietly. We track a few simple signals in production:
- The share of photos rejected by the quality check, by device model.
- The distribution of confidence scores over time.
- Cases where a human reviewer disagreed with the model.
When any of these move, we look at fresh samples, relabel where needed and retrain. It's routine work, but it's what keeps the feature trustworthy.
6. Be clear about what the output means
Computer vision can say a lot, and users will read even more into it. Be precise about what the feature does and doesn't do, show confidence where it helps and keep a person in the loop for decisions that matter. HAI, for example, is designed for general wellness and education and does not diagnose; the report is something to talk through with a professional.
Thinking about a vision feature?
Identity checks, inspection, document capture or something new: we can help you scope it, test it on your real images and ship it into your app. Talk to our AI team →
