Keep your AI features fast and affordable as usage grows

I help Seed through Series B SaaS and AdTech founders decide what to fix when an AI feature gains users and starts costing too much or taking too long to respond.

Selected work

  1. Took Ad Product Collections from launch to $4B+ in annual ad revenue

    Shoppers see these collections in Amazon search in widgets like New Arrivals and Frequently Viewed. Each search needed its own selection of products. I designed how Amazon chose, ranked, and displayed them across the US, UK, and Japan as the program grew to handle more than 100,000 requests a second.

  2. Handled 4.5 million lookups a second. 99% took four milliseconds or less.

    Before Amazon could choose which ads to show, it had to look up details on thousands of products. Those lookups were slowing down the decision. I built the system that put those details within reach while the search was still underway.

  3. Used larger prediction models within a 155-millisecond deadline

    Running a larger model on every search took too long and cost too much. I designed a system that reused a prediction when one was available and refreshed it in real time when it was missing. That let the team use the larger models and still meet the deadline for showing an ad.

How we’d start

Two-week AI & Infrastructure Readiness Audit · fixed fee

Bring one AI feature, your architecture diagrams, and your model and cloud costs. We’ll trace a customer request from start to finish. If the feature recommends an action, we’ll also check what evidence supports the answer.

In a four-hour working session, your engineering lead and I will weigh the options. At the end of two weeks, you’ll have a ranked list of fixes, the trade-offs behind each one, and a 90-day plan your team can execute.

Does your AI feature take too long to respond or cost too much to run? Tell me what the feature does and what you’re seeing. I’ll reply within 48 hours with the first two things I’d check.