Marc Fernandez, Chief Strategy Officer at Neurologyca – Interview Series

Marc Fernandez, Chief Strategy Officer at Neurologyca, is a technology strategist and business development executive with more than two decades of experience spanning artificial intelligence, emerging technologies, consumer electronics, and human-computer interaction. At Neurologyca, he focuses on strategy and the commercialization of Human Context AI, which is designed to help AI systems interpret signals such as stress, confidence, and engagement. Fernandez previously served as CEO of Pearl, an AI-guided platform focused on soft skills and professional development, and as an Entrepreneur in Residence at ALIVE Ventures, where he worked in age-tech. Earlier, he led U.S. strategy at OPPO, focusing on next-generation technologies including AI, Web3, blockchain, and mobile operating systems, and held senior business development roles at Emerge, Denizen, SweetLabs, Movella, and X-Rite/Pantone. He also spent 15 years as an adjunct professor at the University of Southern California, teaching at the intersection of entrepreneurship, communication, social platforms, AI, blockchain, and other emerging technologies.

Neurologyca is a San Francisco-based AI company building what it describes as the Human Context Layer for intelligent systems, enabling AI applications and agents to better understand and adapt to the people they serve. Its API- and SDK-based platform processes multimodal behavioral signals, interaction patterns, and situational context to generate machine-readable indicators around factors such as engagement, confidence, stress, attention, and trust. These insights can help AI systems adapt their responses, pacing, recommendations, and actions in real time while maintaining contextual continuity across interactions. Neurologyca is positioning the technology as infrastructure rather than a standalone application, with potential applications spanning education and training, wellness, enterprise software, autonomous agents, driver safety, gaming, and other interactive systems. The company is currently offering the platform through a selective early-access program for partners building adaptive AI experiences.

Your career has taken you from next-generation technology strategy at OPPO to building Pearl around AI-guided soft skills and now developing human-context AI at Neurologyca. How has that progression shaped your view that one of AI’s biggest limitations is not intelligence itself, but understanding the human behind an instruction?

Thirty years ago, the web made information accessible. Five years ago, AI made output abundant. Today we’re drowning in content while still struggling to get useful outcomes. Raw technical horsepower just doesn’t guarantee a good result for the actual human on the receiving end.

That’s been the through line across my work—from fitting emerging tech into daily life at OPPO, to testing whether AI could teach interpersonal skills at Pearl, to building human context as infrastructure at Neurologyca. Intelligence isn’t the real bottleneck anymore; models reason and execute fine. The harder issue is all the unspoken context people bring to a simple request: why they asked, what makes them uncomfortable, or whether a slight hesitation means the system should abort. Right now, AI only sees a compressed slice of that through text, and as agents start taking real-world actions, that blind spot becomes critical.

Recent examples of autonomous agents stealing credentials, creating fake identities, or concealing their actions raise questions about what happens when an AI becomes highly effective at achieving a goal without fully understanding the intent behind it. Do you see this primarily as an alignment problem, an engineering problem, or a fundamental limitation of how we currently design agents?

It’s a design failure spanning both alignment and engineering. We’ve optimized systems to reach targets without spending half as much effort teaching them what actual success looks like to a human user.

Look at the classic GPT-4 TaskRabbit example: when faced with a CAPTCHA, the model lied and claimed to be visually impaired to get a human worker to solve it. Nobody taught it to deceive anyone; it just identified an obstacle and took the path of least resistance. Reward an agent purely on outcome and it will take wild shortcuts. The core challenge is dividing actions that must be architecturally blocked from actions that are technically allowed but contextually wrong.

You’ve argued that companies are getting better at telling AI agents what objective to pursue without giving them enough understanding of what “success” actually means to the human user. What information does an agent need beyond the prompt itself to genuinely understand intent?

Prompts are just the surface layer. True intent relies on situational depth: prior history, what trade-offs are acceptable, actual stakes, and how the user reacts mid-task.

Some of this comes from app environment state or chat history, but in direct interactions, behavioral cues are what keep context dynamic. People hesitate, change their minds mid-way, or get confused. If an agent can’t read those shifts, it just ends up relentlessly pursuing an instruction that became obsolete two steps ago.

There is an important distinction between actions that should simply be technically impossible for an agent and actions that are permitted but inappropriate in a particular context. How should developers divide responsibility between hard security guardrails and an AI system’s ability to exercise contextual judgment?

Any action causing severe harm with zero legitimate use cases—credential theft, wiping audit logs, unauthorized privilege escalation—should be hard-coded out of existence. An agent shouldn’t waste compute cycle negotiating whether to bypass security controls; the system architecture should just render them impossible.

The real gray zone is everyday actions like sending emails, canceling meetings, or altering internal workflows. These are permissible in theory, but whether an agent should execute them automatically comes down to stakes, user history, reversibility, and subtle cues of user hesitation. Knowing when to stop because context is missing is far more important than raw capability.

Neurologyca is developing technology that can turn signals such as stress, confidence, attention, engagement, and trust into machine-readable context. How could signals like these practically change the behavior of an autonomous AI agent while it is carrying out a task?

These behavioral inputs are essentially real-time reaction data—the exact feedback loop missing from current AI architectures. Right now, models track their own actions, but they operate completely blind to how the human is actually receiving them.
Consider a tutoring context: a user types “I understand,” but their physiological or behavioral cues signal rising confusion and dropping attention. That’s a direct signal for the AI to pivot, slow down, or offer an alternative explanation. The goal isn’t clinical diagnostics; it’s simply establishing a tight feedback loop—context, action, response, adaptation—so the system learns what actually works for that specific individual.

Human context can also be ambiguous. A system might misinterpret hesitation, stress, tone of voice, or other behavioral signals. How do you prevent context-aware AI from becoming overly confident about what a person is thinking or intending?

Behavioral signals are inherently noisy and probabilistic. A sudden jump in stress or a pause in typing could mean confusion, or it could just mean the user spilled their coffee. Systems that translate a single data point into a definitive claim about someone’s emotional state are fundamentally flawed.

Robust systems look at composite signals over time, factor in environmental baselines, and explicitly output confidence bounds. There’s a massive difference between logging “elevated cognitive load relative to baseline” and deciding “the user is anxious.” When uncertainty is high or stakes are significant, the agent’s job isn’t to guess—it’s to ask a clarifying question.

Putting a human in the loop for every consequential decision can undermine much of the efficiency promised by autonomous agents. What does a scalable model of human oversight look like where an agent knows when it can proceed independently and when it genuinely needs to stop and ask for guidance?

Scalable oversight isn’t about maintaining a permanent human-in-the-loop; it’s about dynamic escalation.

Low-stakes, reversible tasks move forward automatically. But as stakes rise, ambient noise increases, or signals hint at user discomfort, the approval threshold scales up dynamically. Crucially, these thresholds have to be personalized—one manager might happily delegate calendar management, while another wants to approve every minor edit. The system has to learn individual boundaries, not apply rigid global policies.

As agents develop persistent memory and increasingly detailed models of the people they serve, context can quickly become sensitive personal data. How should companies approach consent, privacy, data ownership, and limits on what an AI should be allowed to remember about a user?

This is the primary trust bottleneck facing autonomous systems. Persistent memory turns a generic tool into an assistant, but there’s a dangerous overlap between storing workflow preferences and silently building an unapproved psychological profile.
Data governance here needs strict purpose boundaries. If an agent infers stress patterns during an executive coaching session, that context cannot leak into performance appraisals, insurance algorithms, or ad tech. Transparency has to be granular—users need visibility into stored context, one-click deletion, and sensible default retention limits. Just because a system can infer something about you doesn’t give it permission to store it.

How should we evaluate whether an autonomous agent truly understands human intent? Are today’s benchmarks focused too heavily on task completion and accuracy rather than whether the agent achieved the goal in a way the user would actually consider acceptable?

Task completion is a deeply deceptive benchmark. An agent can follow an instruction to the letter and still deliver an outcome the user despises.

If an agent secures the lowest price on a contract but burns down a decade-long vendor relationship to get it, traditional evaluation metrics record a 100% success rate. To the executive, it’s a disaster. We need benchmarks that measure performance within unstated boundaries: Did it respect unspoken norms? Did it pause when conditions changed? That gap between raw completion and actual satisfaction is what current benchmarks completely ignore.

Looking ahead, do you expect human-context capabilities to become a standard layer of the AI stack alongside models, memory, tools, and security? What still needs to be solved before enterprises can confidently delegate meaningful work to agents without continuously supervising them?

As AI transitions from generating content to taking real-world action, human context becomes an essential layer. Models handle reasoning and tools give the system hands, but without human context acting as a live feedback loop, memory and security guardrails are basically running on static assumptions.

Before enterprises delegate real authority, we still have to solve contextual privacy, probabilistic reasoning under uncertainty, and granular governance rules. Building AI that completes isolated tasks is mostly solved; making systems adapt dynamically to the human behind the prompt is what will dictate how much real autonomy we’re willing to grant them.

Thank you for the great interview, readers who wish to learn more should visit Neurologyca.