At Oracle OCI, I work across foundation-model training, fine-tuning, and serving with a focus on doing more with less compute. I led a cross-team multimodal efficiency initiative that cut latency 4x while running on 5x fewer GPUs.
I built an FP8 vLLM checkpoint that let one GB300 replace two at comparable throughput, with no measurable quality regression. I also drove graph compilation, CUDA graph capture, and kernel fusion to reduce inference overhead.
For speech research, I adapted Whisper for variable-length inputs, removing padding compute and deploying the work in OCI's real-time transcription service. I’ve also worked on structured pruning, distillation, and low-resource Hindi-English speech training.
Earlier, at ViaDialog, I trained a French large-vocabulary ASR model from scratch, and at Zaion I developed a Spoken Language Understanding system and contributed architectures to SpeechBrain. My work has appeared at ICASSP and Interspeech.
