Manish Karney

Work / 03 · 2024–2026

Steering YouTube mobile's next decade

YouTube's mobile apps serve more than two billion people on an architecture that had grown organically since 2016. I set the strategy for its first major modernization — and won the alignment, across hundreds of engineers and four apps, to actually execute it.

2B+

users on YouTube mobile apps

4

apps on one platform: Main, Music, Kids, Studio

5%

fewer slow app starts, already delivered

2,016

the last refit before the one I authored

Context

Success is the best camouflage for decay

Nothing was on fire, which was the problem. The apps shipped, the metrics held, and meanwhile the foundation quietly taxed every team: a custom navigation stack the operating system couldn't see, core components forked over and over across apps, and logging too brittle to reliably answer how people actually moved through the product.

Platform decay never loses a quarterly prioritization fight. So I made the long-term cost legible: a one-pager tracing how navigation, forked components, and logging debt compounded into slower features and unattributable regressions, then a summit that turned a diagnosis into a roadmap with names on it.

You don't win platform investment by predicting collapse. You win it by pricing the decay.

The strategy

Sequence pragmatically; say what you won't do

The roadmap deliberately resisted the rewrite fantasy. First: standardized navigation and automated logging — the changes that pay every team immediately and generate the trust to go further. Next: a shared app shell unifying how surfaces compose. The north star — a hybrid, server-driven UI architecture — stayed explicitly aspirational, with non-goals written down so partner teams knew what we would not destabilize.

The operating model was influence, not authority: a small core team setting standards that the Main app, Music, Kids, and Studio adopted because the standards made their work faster. Adoption was the metric. Annual summits kept the verticals aligned and the roadmap honest.

Force multipliers

LLM-accelerated migration, and trust work that never stops

Modernization at this scale meets a brutal arithmetic: migrations measured in engineer-decades. I proposed and lead an LLM-powered answer — Gemini-based agents executing well-specified migration patterns under human review — targeting a 60–70% cost reduction on the largest migration on the roadmap. The lesson generalizes: AI agents are spectacular at executing a crisply defined contract, and the leadership work is making the contract crisp.

AI executes the contract; people own the judgment — ~60–70% less effort on the largest migration

Fig. 01LLM-accelerated migration, gated by human review

automatedhuman gate

The early wins are concrete: slow app starts down 5%, a simplified network stack built on Envoy Mobile with bi-directional streaming, and account-scoping work that keeps child, teen, and private-session data strictly separated in apps used by billions.