STiR coffee and tea magazine

34 STiR coffee and tea | 2025 Quarter 4 Winter (Dec/Jan/Feb) Edwin Calla Durandal, r&d in-charge at Neuhaus Neotec. Photo credit: Neuhaus Neotec The NaNi dashboard from Neuhaus Neotec optimizes machine performance as data accumulates. Photo credit: Neuhaus Neotec ProfilePrint's performance metrics prove that the more the system is used, the more accurate it becomes. Photo credit: ProfilePrint AI Adapted for Coffee's Geography The steep slopes and extreme elevations characterizing quality coffee regions cre- ate unique operational challenges. Traditional mechanization struggles on mountain slopes, and data collection becomes physically demanding. New-age AI-powered solutions adapt systematically to these geographic realities. "Portable, handheld devices overcome terrain limitations entirely," Lai explains. "This equipment can be used directly in fields and processing facilities regardless of terrain." ProfilePrint's analyzers work independently of elevation or slope, relying on analyzing the chemical composition of samples rather than capturing images. Built-in sensors and calibration account for temperature variations across diverse climates. Durandal describes how Neuhaus Neotec incorporates elevation data into their AI models: "Our algorithms have the capacity to adapt and learn from the environ- ment. As time goes by, performance becomes more optimal." The company programs algorithms to consider historical machine behavior across varying conditions, allowing systems to optimize performance as environmental data accumulates. Building Trust in AI Systems Concerns about the reliability of AI in coffee production are legitimate. System perfor- mance depends substantially on data quality and local conditions. Yet, emerging evi- dence suggests that these systems improve markedly with continued use. Many systems learn and improve over time, resulting in increasingly accurate predictions and rec- ommendations with continued use. Qual- ity analysis solutions examining individual beans demonstrate exceptional reliability as datasets expand, although combining AI recommendations with farmer exper- tise yields optimal results. ProfilePrint's performance metrics powerfully illuminate this trajectory. Lai describes their Specialty Arabica Green Bean model's remarkable improvement: "In March 2023, with just 490 samples, the mean absolute error (MAE) was 3.57 points. By July 2025, with 28,000 samples, that error has dropped to 1.07 points." This accuracy falls well within industry requirements for expert cupping assessment. Neuhaus Neotec emphasizes human partnership in algorithm development. "We consider that the experts and opera- tors must be part of the loop, so our algo- rithms are programmed to consider the human-in-the-loop," says Durandal. This approach ensures that operational knowl- edge and contextual understanding shape system behavior rather than operating in- dependently from human expertise. From Skepticism to Strategic Adoption Producer feedback about AI adoption has evolved from cautious skepticism toward enthusiastic integration. This shift reflects real productivity gains and improved decision-making capabilities demonstrated in operational settings.

• Cover