Case study
Two founders had a hypothesis, a pilot grower and no working model. In twelve weeks we built the labelling workflow, trained the vision model, and shipped the web app that turned the greenhouse itself into the demo.
Named only with client approval Figures marked are illustrative Reply within one business day
Vision - line 3 camera
Warehouse detector v4 - 640x640 - centre aisle
38ms
Inference
24/s
Frames read
1,284
Objects tracked
A two-founder agritech startup in the Fraser Valley believed early plant stress could be seen on ordinary cameras before growers noticed it, but had no model, no interface and nothing an investor could hold.OlDevs built a tablet labelling workflow the pilot grower's staff used during their normal walk-throughs, trained a computer vision model on the resulting imagery, and shipped a web app showing stress as a heat map by bay with a daily digest of what changed. The system flagged stress an average of 5 days ahead of manual scouting at 0.91 precision on held-out pilot data. The founders own the model weights, training pipeline and code outright, and used the working pilot to close their seed round.
Key facts
The challenge
The founders were confident that early signs of stress in greenhouse crops show up on camera days before a person spots them walking the rows, and they had a pilot grower prepared to test it. What they did not have was a trained model, an interface a grower would use, or a running system a seed investor could put their hands on. Every conversation stalled at the same point: does it actually work in a real greenhouse.
The hypothesis rested on images nobody had ever labelled. Stress shows up as subtle colour and texture change, and without a consistent way for greenhouse staff to mark what they saw, there was no training set and no honest way to measure whether a model was right.
The grower's staff walk the bays on a fixed schedule and could not absorb a separate data-entry job. Any labelling step that pulled them off task would have been abandoned within a week, taking the training set and the pilot goodwill with it.
A correct prediction is worth nothing as a probability in a log file. Growers needed to know which bay, how urgent, and what had changed since yesterday, in a form they could read between jobs rather than a dashboard they had to learn first.
Seed conversations kept returning to whether the idea survived contact with a working greenhouse. Slides describing the approach could not answer that, and the founders had no running system, no accuracy figure on unseen data and no grower ready to speak to it.
What we built
We began with the labelling workflow rather than the model: a single-purpose tablet tool the pilot grower's staff could use during their existing walk-throughs, which produced a usable training set in three weeks. On top of it we trained a computer vision model to flag stress indicators from fixed camera feeds, and shipped it inside a web app that shows a heat map by bay and a daily digest of what changed. Weekly demos ran in the greenhouse rather than a boardroom, so every improvement was judged against what the growers were seeing that morning.
A focused tablet tool the grower's staff used during their normal rounds: photograph the bay, tag what they saw, carry on. It produced a consistently labelled training set in three weeks without adding a separate pass through the greenhouse for anyone.
A computer vision model trained on the greenhouse's own imagery, tuned for the light, glare and leaf occlusion of a working bay rather than clean reference photographs. It reached 0.91 precision on held-out pilot data.
A web app that renders stress by bay as a heat map and shows a daily digest of what changed. Growers open it, see which bays moved, and walk straight there instead of interpreting model output or filtering a table.
Reproducible training, evaluation and export scripts in the founders' own repository, with labelled data, model weights, camera configuration and deployment notes handed over in full. They can retrain for a new crop or a new grower without calling us.
Process
Week 1: In the greenhouse
We spent the first week walking bays with the pilot grower's staff, learning how scouting actually works and where fixed cameras could sit without getting in the way of the job.
Weeks 1 to 3: Labelling first
The tablet labelling tool shipped before any model work, so data collection started immediately. We revised the tag vocabulary based on what staff reported seeing rather than what we expected.
Weeks 4 to 7: Training and honest evaluation
Training ran against the growing label set with evaluation held out on bays the model had never seen. Precision was reported at every demo, whether it had moved up or down that week.
Weeks 6 to 10: The grower app
Heat map and daily digest were built alongside training, so growers were reacting to real flagged bays in the greenhouse while the model was still improving week to week.
Weeks 11 to 12: Handover and pilot proof
We packaged the pipeline, weights and deployment notes for the founders, then helped them document the pilot honestly for the seed conversations that followed.
Stack
Model and training
Data and labelling
Grower web app
Capture in the greenhouse
Handover
5 days
Average lead time on stress detection versus manual scouting
0.91
Precision on held-out pilot data
3 weeks
From first camera install to first flagged bay
Outcomes
Flagged bays gave the pilot grower an average of 5 days of lead time on stress detection versus manual scouting, which is the difference between adjusting conditions and writing off a crop section.
Evaluation on held-out pilot data returned 0.91 precision, measured on bays the model had never seen. The founders could quote a figure and explain exactly how it was produced.
It took 3 weeks from first camera install to first flagged bay, which meant the grower saw value while the engagement was still running rather than months after handover.
The founders own the model weights, training pipeline, app code and data outright, and took investors to the greenhouse rather than to a slide deck. The seed round closed on the strength of the pilot.
In their words
Investors did not ask to see a deck. They asked to see the greenhouse.
Co-founder & CEO, agritech startup
Agriculture technology
FAQ
Yes, and that is usually where the work begins. On this project the labelling workflow shipped before any model training, because a tool the client's own staff will actually use during their normal day is what produces a training set. Expect the first few weeks to be about capture, tagging and agreeing what a label means.
It depends on how quickly labelled data accumulates and how distinct the signal is. In this engagement the first bay was flagged three weeks after the first camera install, with accuracy improving after that. We show a working demo every week, so you can see the trajectory rather than waiting for a reveal.
You do, in every case. Clients own all code, designs, model weights, training data, accounts and IP produced during the engagement. On this project the founders left with a reproducible training pipeline in their own repository so they could retrain for new crops without coming back to us.
Not without work. Camera placement, crop variety, lighting and growing practice all shift the data, so a model tuned to one site is a starting point rather than a finished product. The reusable part is the pipeline: labelling workflow, training and evaluation scripts, and the app that presents results to growers.
We build the system and the evidence around it, including held-out evaluation and clear reporting on what was measured and how. We do not write your pitch, but a running pilot with defensible numbers tends to carry the conversation. Request a quote and we will reply within one business day.
Let’s connect
Tell us what is not working. We’ll reply within one business day with how we would approach it and a tailored quote.