May 2023 – Present · 3 yrs 1 mo · Washington, United States

• Led optimization of Meta AI Assistant for 2B+ users by advancing transformer inference with PyTorch 2.1, TorchScript, and FBGEMM, achieving a 37% latency reduction and boosting daily engagement by 22%.
• Optimized Meta AI Assistant for 2B+ users by advancing transformer inference with PyTorch 2.1, TorchScript, and FBGEMM, reducing latency 37% and boosting daily engagement 22%.
• Architected retrieval-augmented generation pipelines with FAISS, BM25, Hugging Face, and LangChain, grounding answers from 50M+ documents and cutting hallucinations 28%.
• Deployed LLaMA-Edge models for Ray-Ban smart glasses using 4-bit quantization, Glow compiler, and TorchScript,
• reducing memory usage 45% while maintaining cross-lingual ASR accuracy.
• Directed development of real-time multilingual ASR + TTS systems with wav2vec-U and quantized HiFi-GAN, enabling seamless voice-first experiences across smart glasses and mobile platforms.
• Orchestrated large-scale distributed training across 2,048 GPUs with DeepSpeed ZeRO-3 and NCCL, sustaining 5.7
• TFLOP/GPU throughput to accelerate iteration cycles of 70B-parameter LLaMA models.
• Automated ingestion and transformation of 12TB+ daily conversational logs using Spark, PyArrow, Hive, and Airflow,
• building privacy-compliant RLHF corpora and integrating feature engineering pipelines in SageMaker, MLflow, Docker,
• and Kubernetes (EKS).
• Delivered production-grade MLOps and deployment pipelines with ArgoCD GitOps, canary rollouts, and
• Prometheus/Grafana monitoring, ensuring 99.98% uptime for global inference APIs.
• Partnered with product and research teams to align LLM training with RLHF, applying A/B testing, drift detection, and human evaluation to validate improvements with 99%+ statistical confidence while implementing responsible AI with SHAP, LIME, and Fairlearn.