BeforeAI
Drag to reveal the automatic detection — 9 assets identified: transformer, concrete pole, fuse switch, MV/LV network, public lighting.
A computer-vision system that gets more accurate with every field inspection — with no data team required on the client side.
BeforeAI
Drag to reveal the automatic detection — 9 assets identified: transformer, concrete pole, fuse switch, MV/LV network, public lighting.
One of the largest electric power utilities in the State of Rio de Janeiro needed to automatically catalogue the equipment present in millions of field photographs — concrete, iron and wooden poles, transformers, medium- and low-voltage networks, fuse switches, identification plates and other critical electrical infrastructure assets.
The manual process was slow, expensive and unscalable. Worse: any conventional AI solution would lose accuracy over time as new equipment and field conditions emerged — requiring constant rework and dependence on external specialists. The challenge was not just to detect. It was to deliver a system that kept learning from its own operation.
Each inspection generates data that feeds back into the model. Operators validate or correct detections, and this feedback automatically triggers a new retraining cycle. The most accurate model is promoted to production without manual intervention.
From archive indexing to assisted annotation, auditable training, automatic promotion of the best model, and an operational feedback loop. A replicable process from data collection to production.
Production REST API and .NET/C# SDK integrated into the utility's Windows systems — with no infrastructure migration. The team started using AI without realising they were dealing with machine learning.
Drag each image to compare the original photo with the automatic detection.
BeforeAI
9 simultaneous detections in an urban environment.
BeforeAI
Medium- and low-voltage network with transformer.
BeforeAI
Concrete pole and public lighting.
BeforeAI
Urban infrastructure with multiple assets.
BeforeAI
Spacer network — urban infrastructure variation.
BeforeAI
Rare class: circuit breaker + identification plate.
BeforeAI
Asset coverage on a public road.
The 17 detected asset classes — from poles and transformers to reclosers and concentrators.
In addition to detecting 17 asset types, the system automatically reads the text on identification plates — extracting power rating, manufacturer and asset tag.
BeforeOCR
License plate "ROMAGNOLI" read automatically by the system.
Each dot represents a pole processed by the AI — covering urban, rural and environmentally protected regions of the State of Rio de Janeiro.
Data pulled directly from production databases in June 2026.
Accuracy jumped from 49.7% to 85.6% — a 72% gain driven by data from the operation itself.
| Version | mAP50 |
|---|---|
| v1 (bootstrap) | 49.7% |
| v4 | 52.7% |
| v5 | 60.6% |
| v5.2 | 83.8% |
| v5.3 (current) | 85.6% |
Our model — System + Method + Tool — is replicable in any sector with volumes of visual or documentary data to process intelligently and continuously.
Talk to a specialist