Perplexity

AI consumer products created entirely new support patterns. What happens when millions of people encounter a technology they don't fully trust yet? There was no playbook. So we built one. Native to the product, from scratch. Every conversation was research. Users were pushing the limits of what the technology could do before it caught up to them. What can this thing actually do? Why isn't this working? Some of those conversations became features that millions of people use today. Including things that were pushed back on internally.You can't design a system for something you haven't felt.

Qualitative feedback at scale is an unsolved problem. Still is. How do you capture sentiment precisely and not just quantitatively? We built multiple iterations of intelligence infrastructure trying to get it right. Started with hand-curated reports. Evolved into automated pipelines, sentiment monitoring across product categories, LLM-generated feedback reports wired directly into engineering workflows. Product teams didn't receive complaint summaries. They received hypotheses backed by patterns, with enough context to act. The system evolved as the product did. That was the point.

Default tooling couldn't reason through edge cases. It prioritized deflection over understanding. So we designed a custom system with Parahelp instead. AI triage paired with human review. The AI handled the predictable. Humans handled the ambiguous. The system learned from both. But how do you actually decide what requires a human? How do you flag edge cases when a closed conversation does not equal a happy customer? It usually isn't.

The team was structured around anticipation. Not reaction. Each member held specializations with direct relationships to PMs and engineers. They flagged what users were experiencing before it escalated. They kept asking the questions that kept everything honest. What's broken? Where are we falling short? How would you fix this if it were yours to fix? New solutions always bring new problems. The work was never about solving something once. It was about designing systems that could keep evolving without one person holding all the answers.