Internship
ByteDance
Backend engineering intern · Agent infrastructure
- Shanghai
On the agent infrastructure team, worked on the backends of the TVLA data-collection platform and the GymHub environment platform.
TVLA
Keeping large-scale collection running reliably
With tens of thousands of collectors and hundreds of millions of records, I mapped capacity and stability bottlenecks across collection, upload, processing, and review, then pushed database scaling, read/write separation, index and cache work, fuller monitoring, and API migration. SLI on the established core path rose from 99% to 99.9%.
On the product side, I built task bundles, task recommendations, quota management, review rules, a trial flow for new collectors, user profiles, device-damage reports, and real-device evaluation, and set up dashboards for scale, quality, efficiency, and activity. That work carried the platform through the first half of the year: 12.21 million new collected records, and 87,700 hours of valid video in total.
I also reworked the IDL layering and project structure, filled in CI, unit tests, and scheduled jobs, and organized task-bundle sync and API services, so later changes and debugging cost less.
GymHub
One interface for different environments
Built, from scratch, an environment platform for training, evaluation, annotation, and synthesis. Defined a shared create, delete, act, and observe interface for virtual machines, phones, and sandboxes, and implemented the runtime, project, quota, key, auth-integration, and audit modules.
Decoupled the underlying resources behind a provider interface, and used declarative scheduling plus a runtime state machine to manage each environment's lifecycle, so adding a new resource costs less.
Designed and implemented a warm cache pool. When the image cache hits, virtual-machine startup drops from 70-100 seconds to under 5 seconds.