Life
Uber & Internationalization
Most of my time at Uber went into Rosetta, the internationalization service behind every translated string in the product. It had the highest read throughput of all 2000+ microservices at Uber — over a million requests per second.
Joining During Hypergrowth
After Opendoor, I joined Uber in 2016, in the middle of its hypergrowth phase. The company was launching in new countries at a pace that is hard to overstate now, and every launch needed the product to work in the local language. Not just static text either — dynamic, context-aware translations, for hundreds of millions of users, across very different audiences: riders, drivers, eaters, restaurants.
Rosetta
I came in as the technical lead for Rosetta, Uber's internationalization service. Rosetta translated all user-facing content across the company — more than 70 countries, 30+ languages. What made the job interesting was the scale: Rosetta had the highest read throughput of all 2000+ microservices at Uber, over a million translation requests per second at peak. Translations sat in the critical path of nearly every screen, so the corpus grew by thousands of strings a week and none of it was allowed to be slow.
The system itself was simple in shape: a Go core service with heavily optimized lookup paths, multi-tiered storage with hot-path caching, a React admin portal for managing translations, client libraries for every language Uber used, and connectors to the translation vendors. The hard part was making it fast and keeping it fast.
The Performance Work
Most of my contributions were performance work: a multi-level caching strategy (in-process, Redis, distributed), client-side caches with efficient invalidation, sharded storage for horizontal scaling, a custom serialization format to shrink the memory footprint, and throttling and circuit breakers so a bad day didn't become a terrible one. We built custom benchmark suites, adaptive rate limiting, instrumentation that could track sub-millisecond latencies, and careful gradual rollouts for risky changes. The result was p99 latency under 10ms while serving the most read traffic in the company.
ML in the Translation Workflow
The part I enjoyed most was bringing machine learning into the translation workflow, working with the Michelangelo team (Uber's ML platform). We built systems that suggested translations for new strings based on context, predicted translation quality so human review went where it actually mattered, ranked alternative translations, and flagged likely quality issues automatically. That work meaningfully cut translation costs while improving quality and getting new languages to market faster.
The Communications Platform
Beyond Rosetta, I led the integration of ML into Uber's communications platform — the system behind every push notification, SMS, and email Uber sent. We used models to personalize message content and timing, pick the right channel for each user, and predict engagement so we could stop sending notifications people would ignore or resent. A/B testing at that scale made the effects easy to see: engagement went up substantially, and we saved real money by simply not sending low-value messages. The same approach made driver incentive messaging noticeably more effective.
What the Scale Taught Me
Leading Rosetta meant managing engineers across backend, frontend, and ML, and coordinating with every Uber product line — Rides, Eats, Freight — plus regional teams who each believed, usually correctly, that their market was special. Global products are full of details that don't show up until you ship them: right-to-left languages, pluralization rules that vary wildly between languages, multi-region deployment and replication, region-specific fallbacks when things break.
I also learned that the technical solution is rarely the whole job. Balancing immediate needs against long-term architecture, advocating for resources, keeping a globally distributed set of stakeholders aligned — that was as much the work as the caching strategy was.
What Uber Left Me With
Uber is where I learned to run a system the business genuinely depends on. Performance at scale takes both systemic thinking and obsessive attention to the critical path, and resilience matters as much as raw speed — systems have to degrade gracefully and recover quickly, because at that volume something is always failing somewhere.
The deeper lesson was about pairing technical excellence with measurable business impact: building things that performed well and visibly moved the company forward. That foundation carried directly into my work at Google and YouTube.