Teams at Cerence Inc.

Lead design and training of large-scale transformer and hybrid foundation models across text and multimodal domains. Drive architecture choices (attention variants, MoE), optimize training stability and optimizers, validate scaling laws and compute/data tradeoffs, design loss and alignment strategies (RLHF, DPO, contrastive), and execute distributed training (FSDP, ZeRO-3, tensor/pipeline parallelism) with mixed precision and checkpointing while partnering with ML systems teams.

Lead design and training of large-scale transformer and hybrid foundation models across text and multimodal domains. Own architecture choices, diagnose large-scale training instabilities, optimize scaling tradeoffs (data, compute, model), design loss and alignment strategies, and execute distributed training (FSDP/ZeRO, tensor/pipeline parallelism) with mixed precision to build in-house foundation model expertise.

Lead development and optimization of high-performance LLM inference pipelines across data center, edge, and embedded platforms. Improve latency and throughput via custom CUDA kernels, quantization, KV-cache and batching strategies, and deploy efficient, production-ready runtimes without external vendor dependency.