Skip to content

What I built

Work

Amazon at high QPS. Rippling under a four-second budget. Cursor now — Grok Bot hit #2.

Spend, then a wider org · Staff → EM → Senior EM → Director

Rippling

Corporate cards. Bill pay. The auth path with about four seconds before a charge dies.

Card authorization · p99 vs a 4s budget

0 — 4000ms

600ms
3500ms
decline
  • p99 after
  • p99 before
  • network decline
Miss the budget and someone doesn't pay for dinner. Or AWS.
  • Joined as Staff. Left as Director. Most of what I can point at is Spend.
  • Auth p99: ~3.5s → ~600ms. Not one trick. A pile of unglamorous ones.
  • Hired for that org, including a new Travel product.
  • Indexes are left-to-right. Miss the leading field and you don't have an index.
  • Planner was wrong 1% of the time. That 1% is your p99.
  • Hourly deploys + just-in-time connections = cold-start machine.
  • A timeout that blocks the rest of the request is only half a timeout.
  • p99

    ~3.5s → ~600ms on auth

  • Budget

    ~4 seconds or the network declines it

  • Product

    Spend without forcing HCM / Payroll

About five years · Senior Lead Software Engineer (L6)

Amazon Ads

Tech lead on measurement and frequency capping. Unrecognized traffic. High QPS.

Unrecognized traffic still has a frequency. Someone has to measure the model anyway.
  • Measured ML models on traffic nobody recognized. Org-wide review.
  • Frequency capping via ML. Then blocklists out to the Ad Exchange. ~100k reads/s, ~80k writes/s.
  • Founding member of that team. Two engineers, then more. Demos, book clubs, shared recipes. Yeah, recipes.
  • Measurement

    Online impact of ML models on unrecognized traffic

  • QPS

    ~100k reads/s · ~80k writes/s

  • Team

    Founding member. Mentoring. Yeah, recipes.