Ruben Sandoval Davila had a working product, a growing user base, and what most founders in digital health would consider a reasonable hypothesis: automate everything. Remove the human bottleneck from clinical nutrition. Let the algorithm handle intake, documentation, meal planning, follow-ups, and billing. Scale the thing until the unit economics make investors weep with joy. The hypothesis was clean. The data that came back was not.
When Sandoval’s team at Avena Health fully automated their platform, patient retention collapsed. The system had been holding 40% of its users for eleven months or longer, a number that would be exceptional for almost any consumer health product. After full automation, 9% of users remained active after one month. After three months, 2%.
Two percent. On a platform that had previously kept four out of ten users engaged for nearly a year. The automation worked exactly as designed. Patients left anyway.
Sandoval is the co-founder and CTO of Avena Technologies, the company behind one of Latin America’s largest digital nutrition platforms. The system he has built has processed more than four million patient consultations across a network of over fifteen thousand health specialists. It is not a beta or an MVP running on conference demo logic. It is a production clinical system, and the architecture that powers it was designed in direct response to the failure described above.
Most founders, confronted with a retention collapse of that magnitude, would have treated it as a product engagement problem. Add gamification. Redesign the onboarding flow. Send more push notifications. Hire a behavioral psychologist to audit the user journey. Sandoval did none of that. He treated it as an architectural problem. The platform had removed human specialists from the care loop to gain efficiency. The efficiency gains were real. The documentation was faster, the follow-ups were more consistent, and the system could handle far more patients per hour. None of that mattered, because the patients stopped showing up.
“The users who stayed on the platform before automation weren’t staying because the technology was better,” Sandoval said. “They were staying because a human being was paying attention to their progress. When we took that away, no amount of personalization from the algorithm could compensate for it.”
The rebuild did not involve removing the automation. It involved redrawing the line between what the algorithm should handle and what a human should handle, and then enforcing that line in the software itself. Sandoval calls the result a hybrid AI-human care model. AI handles the clinical tasks that practitioners do slowly and inconsistently: documentation, plan generation, follow-up scheduling, billing. Human specialists handle the tasks that AI does poorly: clinical judgment, rapport, and the moments where a patient needs to feel heard rather than processed.
The distinction is easy to state. The engineering is not. The architecture underneath the model has three layers that Sandoval designed and built over the course of several years.
The first is what the team calls the End-to-End Clinical Loop. Every consultation generates structured data that feeds forward into the next recommendation, follow-up, and billing event. Consult, structure, act, track, re-engage. Nothing is re-entered manually at any step. The loop is closed by design rather than by practitioner discipline, which means the data quality does not degrade as the system scales.
The second is a Constraint-Based Nutrition Engine. Clinical nutrition involves hard constraints: macronutrient targets, drug interactions, allergies, chronic conditions, religious dietary restrictions. Most AI systems in this space generate outputs and leave the practitioner to validate them against these constraints manually. Sandoval’s engine enforces them at the data layer. The system cannot produce a meal plan that violates a patient’s clinical parameters. The constraint is structural, not advisory.
The third layer is the workflow integration itself. Rather than building AI features and presenting them to the practitioner as tools to opt into, Sandoval embedded the AI at each step in the clinical process. The practitioner does not choose when to invoke the algorithm. The system determines, based on the data available at that step, whether automation or human action will produce the better result. The effect is a reduction in cognitive load on the specialist, which is a measurable variable in clinical software adoption and one that most health tech companies treat as someone else’s problem.
The retention numbers came back. Forty percent of patients now stay on the platform for eleven months or longer, restored to the rate the system achieved before the automation experiment. That number is not a marketing figure cited in a pitch deck. It is the empirical result of an architectural decision made in response to a specific, measured failure.
The clinical implications are significant. A nutrition platform that retains patients for nearly a year produces fundamentally different health outcomes than one that loses them in weeks. Behavioral change in nutrition is a long-cycle process. The patients who stay for eleven months are the ones whose dietary habits actually shift. The ones who leave after three months are the ones who downloaded an app and forgot about it.
In September 2022, Avena was selected for the Google for Startups Accelerator for Latino Founders, a cohort confirmed on Google’s official blog. The selection criteria emphasized demonstrated traction and deep AI/ML technical expertise. By the time of that selection, Sandoval had already scaled the platform well beyond what most accelerator companies reach by graduation.
A caveat is worth noting. The retention data is self-reported. Avena has not published these figures in a peer-reviewed study or submitted them to an independent audit. The numbers carry the credibility of a founder willing to describe his own product’s failure in specific terms, which is unusual but not the same as external verification. Sandoval says he would welcome an independent review. For now, the four-million-consultation scale of the platform and the enterprise clients who use it serve as indirect validators.
The most interesting thing about this story, assuming the data holds, is what it suggests about the broader industry. The prevailing pitch in digital health is full automation, humans out of the loop, AI handling everything. Sandoval had that product. He watched it fail on the metric that determines whether a health platform works. The architecture he built in response treats human attention not as an overhead cost to be eliminated but as a system component to be deployed where it produces the most value. That finding did not come from a research paper. It came from production data, the kind you only get by building the wrong thing first and being honest about what the numbers say.
