Mahib Hosain

Google on-device
LLM gateway

Aratea

100+ clients onboarded onto the on-device LLM gateway

One of two engineers building a gateway with a common API for image and text generation through Google’s Imagen and Bard models. It supported Android AI Wallpaper, Android Auto, Messages, and other experiences. I owned feature prioritization, design, and implementation, as well as metrics, quotas, experimentation, CI/CD, and on-call.

Powered experiences demonstrated at Google I/O 2023 and showcased at CES.

Automatic language version upgrades at Stripe

From quarterly migration projects to continuous language upgrades

I led Stripe’s automated JDK upgrade platform: using live Bazel dependency graphs to determine eligibility, partition reviewable changes, and surface blockers.

Stripe context: 26M lines of JVM code · 700+ Java services

LLM-aware
Bazel output

--config=llm

Less build noise in agent context, lower token costs

I implemented an LLM-aware Bazel output filter across Stripe’s polyglot monorepo—Ruby, JVM/Python, and Go. Enabled with --config=llm, it strips ANSI escape codes and verbose build/test noise before the output reaches coding-agent context windows.

The filter reduces token costs for Stripe’s Minions—built on a fork of Goose—as well as Codex and Claude Code.

AI assisted code
translation at scale

Removing the Scala constraints keeping Stripe’s JVM codebase on legacy versions

I built llmigrate, a Go CLI orchestrating five specialized Claude agents to translate Scala into Java, repair Bazel builds, translate tests, and check JVM compatibility.

Generated bridge code and cross-language tests validate Java implementations against Scala behavior.

On-device data rollouts

ML models, font files, and other data packages across Google apps

Before Aratea, I built telemetry and analytics for measuring data-package rollouts on devices across Google apps, including Search and YouTube.

Built pipelines processing 1B+ device events daily, aggregating and visualizing rollout metrics with HLL / KLL sketches. Configured anomaly detection and automated alerts to catch rollout regressions, and integrated rollouts with A/B testing.

Bazel performance

Remote execution, caching, and dependency hygiene across 700+ Java services

Improved remote execution for JVM development, enforced versioned dependencies in Starlark, and introduced Java quality gates with Error Prone.