Cyril Tamraz

Wake Word Detection:

• Tech lead for the integration of the wake word detection feature of Alexa across automotive partners.
• Strong experience with building audio pipelines, webRTC Voice Activity Detection, keyword spotting.

Edge Machine Learning:

• Main contributor to an inference engine designed in C for optimized latency and portability across several computer architectures, including ARM and Qualcomm DSPs. This engine is currently used for Alexa wake word detection, acoustic event detection, and on-device speech generation.
• Optimized the serialization of TensorFlow and PyTorch models to run on the edge inference engine.
• Designed a custom RNN operator in C that uses graph optimizations to accelerate the deployment of a series of parallel LSTMs, achieving a 30% reduction in MIPS on average across all supported chips.
• Wrote assembly code to optimize convolutions and other neural network operators on a Qualcomm QCC5171 chip, achieving a 50% reduction in MIPS on that particular hardware.

Infrastructure

• Main contributor to the Jenkins test infrastructure for the Alexa edge wake word detection feature, which is CICD across over 200 computer architectures or emulators scaled in EC2 instances.

• Implemented a Windows service with C# (NET Framework) to dial input phone numbers, capture call recordings, and transcribe audio via the Google Cloud Speech-to-Text API.
• Developed a rule engine for classifying numbers based on the transcription. Enhanced pattern-matching efficiency with a rolling KMP algorithm, achieving a 70% reduction in processing time.
• Built and deployed AWS microservices, including Lambda, to monitor vendors' S3 buckets, compress and store the data in the company's bucket or internal databases to reduce storage costs by 20%.

• Clustered patents that are semantically similar using NLTK, and predicted patent activity for DeepMind and its competitors using vector autoregression (VAR) on data from the Google Patent database.
• Undertook research in optimization for statistical learning. Studied the relations between first-order algorithms and data conditioning, and the usage of relative continuity to solve non-Lipschitz problems.