The Challenge
PoolWater Pro’s business runs on one repeated action: testing water chemistry. Every pool serviced means a test strip dipped, read, and translated into a treatment decision. The company had two ways to do it, and both were broken.
The accurate way was the lab. Send samples out, wait 48 hours for results, then act on chemistry that was two days stale. Water conditions change daily; a 48-hour-old reading is a report on the past, not a basis for treatment. Customers waited, technicians made return trips, and the operational cost of every test compounded across the service fleet.
The fast way was human eyeballs. Technicians read strip colors in the field and judged them against a reference chart — in direct sunlight, in shade, on strips from different batches. The result was a 23% human error rate in strip readings. Nearly one in four field tests produced a wrong value, which meant wrong chemical dosing, callbacks, and in the worst cases, unsafe water that a paying customer swam in.
PoolWater Pro needed lab-grade accuracy at field speed, on hardware every technician already carried: a phone.
Our Solution
OmniMinds built a computer vision system that turns any smartphone camera into a water chemistry lab.
A YOLOv8 model trained on 50,000+ test-strip images. We assembled and curated a training set covering the real conditions strips are read in — varied lighting, angles, strip batches, and degradation states. The model was trained in PyTorch on YOLOv8, chosen for its accuracy-to-latency profile on mobile-class hardware.
Real-time detection of 7 mineral parameters. A technician points the phone at a dipped strip; the model locates each reagent pad, corrects for lighting, and reads all seven mineral parameters simultaneously. No reference chart, no squinting, no judgment calls. The result renders on screen with the treatment-relevant values in 15 seconds.
A nightly retraining pipeline. This is where the DevOps practice earned its place in the engagement. Every field reading — including the edge cases the model is least confident about — feeds an MLOps pipeline that retrains the model nightly. The system PoolWater Pro runs today is measurably better than the one we shipped, and it improves on its own schedule without an engineer in the loop. Model versioning, evaluation gates, and automatic rollback are built in, so a bad training run can never reach a technician’s phone.
The senior engineers in the pod designed the model architecture, the labeling strategy, and the evaluation gates; AI agents accelerated dataset preparation, augmentation, and pipeline scaffolding. That division of labor is standard in every OmniMinds pod, and it is why the timeline stayed short.
The Results
- 15-second analysis, down from 48 hours. The lab wait is gone. Chemistry results are now available before the technician has put the strip down — a 95% reduction in end-to-end testing time.
- 98% accuracy. The 23% human error rate in field readings dropped to a 2% model error rate, with low-confidence readings flagged for a second look rather than silently guessed.
- $210K saved per year. Lab fees, repeat visits, wasted chemicals from misdosing, and callback labor — eliminated or sharply reduced across the fleet.
The second-order effects are the ones PoolWater Pro’s leadership talks about. Technicians complete more stops per day because every visit resolves on the spot. Customers get treatment decisions during the appointment instead of a follow-up call two days later. And every reading captured in the field makes the model better — the company now owns a proprietary, compounding data asset in an industry that still reads colors off a paper chart.
Why It Worked
PoolWater Pro did not need a machine learning research project; they needed a business result — accurate readings, in the field, now — and that is how the engagement was scoped. Fixed outcomes: accuracy threshold, latency target, cost savings. Not a burn rate.
The AI-augmented pod model fit this problem unusually well. Building a production vision system is 20% modeling and 80% everything around it: dataset quality, evaluation rigor, mobile deployment, and the retraining loop that keeps accuracy from decaying in the real world. OmniMinds’ senior engineers — Top 1% Expert-Vetted, with a 100% Job Success record across 34+ projects — owned that unglamorous 80%, while AI agents compressed the schedule on data preparation and pipeline work.
Most vendors would have shipped the model and left. We shipped the system that keeps the model good — because the outcome PoolWater Pro bought was not “a model.” It was 98% accuracy, every day, indefinitely.