Thad Hughes Email & Phone Number

San Francisco, California, Us

We build AI systems that can reason.

Mountain View, Ca, Us

Our team built cutting-edge computer vision systems for medical imaging applications. I made multiple trips to Chicago to collaborate with doctors at Northwestern, learning about their systems and product requirements and bringing insights back to the team. I researched and developed ML models to detect tuberculosis in chest X-rays, locate misplaced medical devices, and identify similar medical images using embedding vectors. I also developed techniques to spot dataset biases and reduce model cheating, and served as team liaison with the medical image labeling team to ensure alignment on product feature roadmaps.

Mountain View, Ca, Us

I served as Tech Lead Manager (TLM) for hotword detection during Google Home’s launch, building Google's first combined “Ok/Hey Google” hotword models and improving performance for accented English and child speakers by ~40%. I also managed a cascade of hotword detection models for Android, including low-power embedded models running on Digital Signal Processing (DSP) chips, and guided hotword data collection and model training across 50 different locales.

Mountain View, Ca, Us

Mountain View, Ca, Us

I led development of German Voice Search, managing data collection in Germany, and training and debugging acoustic and language models, the German pronunciation lexicon, and text normalization.I dramatically improved ASR and hotword detection in far-field and noise by developing novel multichannel techniques for background noise suppression, blending ML trainable parameters into traditional signal processing algorithms, and realistically augmenting audio training data, reducing Word Error Rate by 60% and hotword False Rejections by 50%.I single-handedly built the DataHound application for Android and published a paper about this pioneering approach to speech data collection. The app was ultimately used to crowd-source recording of 1,000s of hours of transcribed speech for acoustic model training.I redesigned the acoustic model training and evaluation pipeline's configuration, dramatically reducing the cost of internationalization.I served as a GOOG-411 software developer and later Tech Lead, where I designed a sophisticated A/B comparison dashboard to detect problems in voice application logic, ASR performance, and telephony by using logs to detect significant differences in event distributions.

Mountain View, Ca, Us

As a Google 20% project, I taught myself Android development by single-handedly authoring QuickDic, an offline, open-source, multilingual dictionary for Android. When I handed it off in 2014, it had 4.4/5 stars and 1M+ downloads, and is still (as of 2023) supported by another developer and available in the Play Store.

Stanford, Ca, Us

I researched pronunciation modeling for ASR using Dynamic Bayesian Networks, and word sense disambiguation and automatic clustering of related word senses, developing a novel graph-based clustering algorithm.

Mountain View, Ca, Us

We built a VoiceXML-based telephone voicemail dialogue system. I added features, introduced unit testing to the project, and earned Google readability in JavaScript, which is rare for an intern.

Us

My team built an optimized order execution engine to trade US equities. I improved the efficiency of market data handling by 15x.I did this by designing and implementing a distributed, multi-threaded market data library to continuously query full-book prices streaming from multiple sources using TIBCO's pub-sub messaging system, and, after profiling, implementing custom data structures to minimize small object allocations. The resulting system handled 15x more data per machine.

Starnberg, Bavaria, De

I taught software development, data structures and algorithms, computer architecture, and digital logic design to high school students, following the International Baccalaureate (IB) curriculum, creating original lessons, exams and and programming assignments, including a Swing UI frontend app to construct photomosaics of a target image from a collection of candidate tile images.I also returned as a guest lecturer to computer science classes in 2013 and 2023, presenting on Machine Learning and Data Science.

Us

M.S., Computer Science

Bachelor'S Degree, Computer Science