Senior Software Engineer at Meta

Contributed to the Reels ranking stack, owning a transformer-based sequential model (TransAct-style) with multi-task heads for watch-time, like, share, and skip; achieved ~3.5% watch-time lift in A/B testing. Co-designed GPU sharding and FSDP training recipes for a DLRM-style ranker with ~20B embedding-table parameters, reducing per-epoch training time by ~30% and keeping P99 inference under 200ms. Owned the video-tower component of the Explore retrieval pipeline with a two-tower setup; improved candidate recall@1000 by ~6% offline. Shipped creator-side moderation features using YOLO and DeepLab integrated via OpenCV preprocessing, feeding downstream safety classifiers with calibrated confidence scores. Prototyped a BERT/RoBERTa + CLIP-style retrieval stack with FAISS for creator-assist features, informing the agent-assist roadmap after LLM maturation.

Senior Applied Scientist at Microsoft Research

Contributed to Project Science Engine, owning the retrieval component that progressed from prototype to pilot in 2 Fortune 500 tenant environments (Pharma and Energy). Built BERT/T5-based NLP services for scientific-literature search, dense-passage retrieval, and summarization; users ran ~25% more follow-up queries per session than on a keyword-search baseline. Ran fine-tuning experiments on domain transformers using DeepSpeed ZeRO with mixed precision on Azure ML, achieving ~8–12% F1 improvement over off-the-shelf BERT on internal life-sciences benchmarks. Contributed research on gene/drug/protein interaction modeling that shipped into production.

Senior Machine Learning Engineer at Schlumberger (SLB)

Co-developed a reinforcement-learning controller for real-time hydraulic-fracturing pumps (VAE pre-training + Dueling DDQN in PyTorch) and built the ML-optimized edge stack for deployment. A 6-month field trial showed 20% efficiency gain and 35% reduction in unplanned downtime, translating to ~$56M/year in operational savings.