Brandon Sanders

A Rust-based Kubernetes microservice platform that enables internal and external Wayfair customers to find products that are visually or semantically similar to one another based on a variety of classification and vector embedding machine learning (ML) models.

Upon launching, this service drove over $5,000,000 per year in new customer conversions by enabling our sales agents to rapidly identify products similar to customer-provided imagery. It also saved over $330,000 in annual cloud costs from legacy search systems and attained 10x+ throughput gains over legacy search systems.

I was the staff software engineer and architect for this platform. I led the initial architecture and foundations for the different microservices and built a novel framework that enabled us to serve models from any framework (TensorFlow, PyTorch, ONNX, etc.) from the same web service with per-inference latencies that outperformed typical TensorFlow Serving solutions. Additionally, I built a novel framework that enabled Rust microservices to invoke data scientists' original Python pre-processing logic with minimal overhead, enabling our ML engineers to develop safe and performant Rust systems and our scientists to use the latest Python AI/ML frameworks.

Built with Java, Kubernetes, Kafka, Beam, and Flink, this globally distributed event-driven platform powers Wayfair's online advertising.

Upon delivery, this platform generated over $300,000 in net new revenue by providing more advanced product feed customization tools for marketing partners and now drives more than $1.5 billion in global attributable revenue.

I was a senior software engineer and architect for this platform, leading the design and implementation of the event-driven distributed data architecture, applying my background in mechanical sympathy to realize 10x+ latency and throughput improvements compared to the legacy feeds platform, and implementing distributed tracing so we could track the lifecycle of a product from its first moments in the catalog to its final ad listing with a vendor.

A Java-based microservice that serves as a low-latency (<10ms) AI/ML feature storage service. On its launch day, this service outperformed the latency of Google's equivalent offerings by nearly 10x.

Upon deployment, this service reduced annual cloud storage costs by $600,000 and reduced the lag between creating features and being servable from seven days to one day.

I was the senior engineer on this microservice, leading its design, development, delivery, and operations. I built the service with a mixture of Java, non-blocking IO, and the LMAX Disruptor framework, enabling the service to maintain low per-request latencies for web clients while elastically scaling to handle bursts of high-throughput requests to support offline batch processing jobs. The service is backed by an Aerospike cache, which is rewarmed with the latest daily features from a wide range of internal feature storage systems.

Researched and developed a pure Rust implementation of the original OpenAI GPT-2 TensorFlow model, including the novel byte-pair encoding tokenizer and temperature-based nucleus sampling. To successfully develop and translate this model to Rust, I had to fully understand how the original GPT-2 model worked in Python and TensorFlow—down to the individual tensors. The Rust implementation is Tensor-for-Tensor compatible with the original Python GPT-2 model, and it can load and run from OpenAI's original training weights for GPT-2.

Leveraging my GPT for the Rust crate, I published an early 2023 article explaining at a high level how generative artificial intelligence, like GPT, can be used to label and classify data without supervision. As part of this article, I developed a full command-line application that leveraged the hidden layers of a GPT-2 model to embed quotes from a fictional universe and predict which culture in that universe a given quote was most likely to originate from. Although GPT-2 is not a strong fit for embedding arbitrary data, and my training data set was relatively small, the resulting KNN-based classifier was approximately 60% accurate. For reference, truly random results would be only approximately 20% correct.