Artificial intelligence is no longer just a distant promise for the aquaculture industry. Today, it is being used in processing lines, at farming facilities, and on analytics platforms. But for it to truly work, the data has to be there first. We spoke with four industry leaders about how they are building that foundation and what comes next.
Ximena Tapia: Generating reliable data is the foundation for applying AI
Ximena Tapia has been working at the intersection of technology and aquaculture production for over a decade. For her, the main bottleneck isn’t AI itself, but the quality of the data that feeds it. “The facilities that can truly benefit from artificial intelligence today are those that have been consistently collecting data for years,” she says.
Tapia’s argument is straightforward: before discussing predictive models or advanced automation, we must ensure that the data being collected is accurate, properly labeled, and representative of the process’s actual variability. “A model trained on bad data will make bad decisions, and in a processing plant, that translates to a specific cost per ton,” he explains.
Without clean, organized, and accessible data, artificial intelligence has nowhere to start.
Ximena Tapia — Aquaculture Traceability Specialist
In this context, solutions like Q-Sort and Smolt Tracker offer value that goes beyond real-time sorting: every structured measurement they generate is a data point that, when accumulated over time, enables the training of more accurate models and the creation of a knowledge base specific to each plant.
Matías Poupin and Tomás Vera: Decades of Using and Implementing AI
Matías Poupin and Tomás Vera represent the perspective of those who have been implementing machine learning algorithms in real-world production environments for years. For them, the debate is not whether AI works, but how it can be reliably deployed in factory settings where hardware is exposed to water, temperatures fluctuate, and operators have varying levels of technical training.
The case of Blue River Technology—acquired by John Deere and known for its computer vision systems applied to agriculture—serves as a direct point of reference for Poupin: "What they did with vegetable crops, we are doing with fish. The implementation logic is similar: robust sensors, models trained using data from the actual operating environment, and short feedback cycles."
It's not enough to have the right model. It needs to run stably, in real time, without relying on external connections.
Matías Poupin — Engineer in Applied AI
Vera adds that one of the most important lessons she has learned over the years is that the model is only one part of the system. The quality of the cameras, the lighting, hardware maintenance, and regular calibration are just as critical to ensuring that the final result is reliable.
Pablo Durán and Jonathan Bello (Akva Group): How to Implement AI in Real-World Environments
At Akva Group, Pablo Durán and Jonathan Bello work directly with the plant teams that implement these technologies. Their approach is more pragmatic: AI must solve specific problems, reduce human error, and free up operators to focus on higher-value tasks, without creating new complexities that the team cannot handle.
"The challenge of implementation isn't technical—it's human," says Durán. "You have to get the field operator to trust the system. If the system makes a mistake and there’s no mechanism to correct it in real time, you lose all the trust you’ve built up.” That’s why the solutions that work best are those that show their reasoning, allow for manual intervention, and learn from cases where the operator disagrees with the model’s decision.
AI doesn't replace people on the ground, but it does open up a world of possibilities that didn't exist before.
Jonathan Bello — Akva Group
Bello also points out that the traceability provided by these tools has a direct impact on process auditing and certification—something that European and North American markets are demanding with increasing rigor. "Having an automatic, structured, and verifiable record of each batch is no longer a competitive advantage; it will become a requirement."
Useful, applied AI focused on value
The common thread running through these discussions is clear: the most valuable artificial intelligence in the salmon industry is not the most sophisticated or the most innovative, but rather the kind that solves a real problem, generates reliable data, and can operate reliably under the conditions of a processing plant.
Lythium operates precisely in this space. Every solution—from Q-Sort to Smolt Tracker—is designed to capture accurate, structured, and actionable data, and to do so autonomously, in real time, and without relying on the cloud for critical operations. That is the foundation. What is built upon it is what customers who have been working with us for years are already achieving.
