"Design for human moments, then let AI scale."

Modern social platforms are discovery engines, and AI/ML now determines how people find content, communities, and answers. In this piece, I outline five guardrails that separate durable progress from short-term spikes. I'm drawing on experience building AI/ML consumer products at scale, with examples from earlier work in consumer fintech. If you build for large audiences, or want to, I hope these lessons help, and I would love to hear your take as you apply them.

Start from Real User Jobs: Design Around Core Human Moments

Effective product design begins by observing how people naturally use a platform and defining core "jobs to be done" around human behaviors rather than technical features. Users arrive with specific, often unspoken goals—to "join" a community (for a new hobby), to "ask" a question (to get instructional advice), or simply to "watch" content (for entertainment or information). Strategic focus is placed on users whom traditional systems often underserve, such as low-signal or low-frequency users. For many of these users, community-driven experiences are a major—sometimes the only—use case; the objective is to identify their specific needs (commerce, niche interests, lifestyle) and show tangible value that helps them graduate to higher-loyalty users. Solutions are designed to address these moments directly. For instance, recommendation systems can be tuned beyond power-user optimization to surface relevant local or interest-based communities for these cohorts, so the system serves real human needs.

Design Ranking Around Human Value, Not Short-Term Metrics

Ranking systems should be built to support the primary product experience. Understand what content delivers the highest value for users and the business, then build objectives and feedback loops around that. Users frequently experience low diversity and perceive some recommendations as too broad; quality is characterized by relevance, diversity, and usefulness. To reason about subjective quality, platforms can rely on aggregated qualitative signals rather than manual heuristics. These signals improve retrieval and ranking in a general sense. Objective functions should emphasize long-term outcomes—engagement that supports sustained use and trust, not just one-off clicks.

Balance Social, Creator, and AI-Driven Content

The key is to safeguard core social value while integrating creator and AI-generated content in proportions that enhance discovery without overwhelming it.

Many consumer platforms are evolving into discovery engines for unconnected content that expand interest-based exploration while still leveraging known preferences. Cannibalization must be actively managed, and signals need to be applied consistently across surfaces. When done carefully, discovery can broaden without displacing the interactions users value most.

Design for Uneven Ecosystems: Speed and Trust

AI/ML systems must deliver fast, trustworthy experiences across diverse conditions—network quality, device constraints, and digital literacy. A persistent challenge is the lag between a new signal (for example, a newly expressed interest) and the system surfacing relevant content; slow adaptation erodes trust. Prioritizing recent behavior helps systems feel responsive without exposing internal mechanisms. The result is responsiveness users can feel and a system they can rely on.

Make Experimentation Safe at Scale

At very large scale, even small negative changes can meaningfully damage trust and outcomes. Measurement must account for interconnected experiences and focus on long-term value, not short-term spikes. Operational guardrails include careful monitoring, qualitative review, and conservative rollout practices. Advances in automated evaluation can help assess quality consistently where manual review does not scale. The art is empathy for high-leverage human moments; the science is AI-enabled systems that can scale those moments responsibly. Together, they turn large consumer platforms from static directories into responsive discovery engines.