New York, New York, United States
Migrated Instagram’s Async Worker Service to a new interface to allow deferred off-peak compute resource utilization with support for argument (de)serialization, function name resolution, resource limiting, and exception handling for 100% of the asynchronous job requests to serve upwards of 30M req/min. Enabled the re-architecture of Async Tier’s submitter and dispatcher components to improve scalability and extensibility through the Async Worker migration. Introduced a simple way to adapt a standard (de)serialization format (Thrift) for asynchronous job arguments so that Privacy Aware Infrastructure (PAI) can consistently validate the semantic representation of data that flows across services at Meta.
Established modern cloud-based development practices by migrating legacy PHP applications to Go micro-services. Revamped Imgur’s infrastructure with Kubernetes to make operations simple, scalable, and fault tolerant. Achieved a 50% increase in lookup speeds for applications across different networks by developing a fast DNS-based service discovery daemon. Built a network check plugin for Kubernetes to ensure uptime and monitor reachability across multiple services. Overhauled Imgur’s image classifier to protect against malicious uploads (decompression bombs, pixel flooding, etc.); tuned models to reduce manual moderation efforts for user-submitted content by 20%. Designed a system with a dynamic pool of sandboxed proxies for fetching remote content uploaded on Imgur’s Video-to- GIF endpoint. Engineered a tool to identify and de-duplicate 442M images with perceptual hashing to reduce storage backend costs. Ensured best security practices through investigations and security fixes for issues reported on Imgur’s bug-bounty program.
Developed a service for dynamic routing, configuration management, and maintenance of Yahoo!’s Video Policy API (VPA) infrastructure to serve upwards of 15K req/s (with 16ms 95th percentile latency) during peak usage. Integrated the Video Policy Engine (VPE) with Linux Virtual Server (LVS) and Keepalived (with VRRPv2) for better load-balancing, failover and resiliency. Implemented a prototype plugin for Apache Traffic Server (ATS) to fetch video streams with Amazon Web Services (AWS) S3 bucket as an origin server. Designed a testing framework to catch error-prone configuration changes to VPE before being deployed to production. Migrated Yahoo!’s transcoding, live-encoding and content delivery infrastructure to a new and improved network stack. Supported Yahoo! NFL stream – the first free, global live-stream of a regular season NFL game which reached over 33.6M total views across all devices on Yahoo! and Tumblr.
Developed features for MILE Lab’s optical character recognition (OCR) engine: a new correction and performance evaluation tool for recognized text; user interface for block-wise overlay display for text on image, and a search feature to highlight matches on the image canvas.