United States
· Designed and built a speaker diarization evaluation platform and internal dashboard, benchmarking models across datasets, metrics, collar values, and reference-label variants; validated by reproducing published results. Built a diarization model on that foundation, achieving state-of-the-art DER. · Improved the quality and reliability of the Pegasus video segmentation pipeline: lifted segmentation F1 via semantic-similarity boundary resolution and cross-chunk merging of split segments; hardened robustness via prompt tuning, entropy-increasing retries, and coverage/validity checks. · Cut Marengo serving costs - up to 27% higher throughput and 43% higher GPU utilization - by replacing Ray Serve's default request router: benchmarked in-house and Anyscale variants, specified the replacement's design for Anyscale to implement, and diagnosed and directed fixes to their scheduling logic through rollout. · Drove Marengo video-embedding pipeline optimizations: cut tensor-read and cross-node-transfer latencies by 40-75% per targeted stage (dataloader/memory rewrites, adaptive frame-rate sampling, multi-connection gRPC tuning); AOT-compiled the scene-boundary-detection models; and right-sized GPU instances. · Designed and led a phased, dual-write migration of the Pinecone index with validation tooling and staged cutover, avoiding disruption to live services; the new smaller-namespace schema cut query costs.
· Boosted search engagement by adding large pin-ID embedding tables to the search ranking model. · Improved efficiency of the search ranking model via sparsification and quantization of large features. · Stood up shadow-traffic serving clusters to power a new serving-cost estimation tool; onboarded the ranking team.
· Designed and implemented edge classification for LinkedIn Sales Navigator. · Tuned relevance-model training labels with Bayesian optimization to lift business metrics that could not be optimized directly. · Designed an explore/exploit approach for Ads Visual Optimization.
· Researched, developed, and deployed plant-based formula generation methods for mimicking animal-based foods, using representation learning (e.g.\ VAEs, set NNs), black-box optimization, and contextual/combinatorial bandits. · Built a Bayesian-optimization design-of-experiments platform used by scientists and chefs to accelerate development of plant-based foods. · Developed novel preprocessing methods for nutritional, physiochemical, and perceptual data sources.
· Researched deep-learning wave-to-wave methods for digital audio processing that mimic hardware audio devices; embedded devices and their settings in an explorable latent space, enabling interpolation and creation of new sounds.