Matt Forbes

Western Washington University

Senior Software Engineer at Grand Rounds, Inc.

Bellevue, Washington

Grand Rounds, Inc.

Senior Software Engineer

Software/ML engineer on the Perception Team. Work ranging from low-level integration of sensors, implementation of real-time perception runtime, performance instrumentation, and batch data processing for deep-learning models.* Built multi-LiDAR synchronization and aggregation component. Processes millions of points per second across multiple sensors. Includes full instrumentation and health monitoring.* Designed and built the object detection stack, including implementation of both geometric-based and deep-learning pipelines. Coordination of asynchronous computation across both CPU and multiple GPUs to achieve consistent latency targets. * Drastically reduced the runtime of the deep-learning inference pipeline by streamlining the entire process to run in a single cuda stream. This included rewriting all pre-processing code to run on GPU using the cuda library.* Implemented the sensor fusion framework used to build multi-modal detection pipelines. * Built out an observability framework used across the entire self-driving system to instrument time-series data. This is used for long-term performance and regression monitoring as well as detailed debugging for individual issues. * Created a batch processing system which ran a sequence of models and heuristics for every single input sample collected in the field, storing the results in persistent queryable storage. This was used to inform the automated processes to select which data should be labeled for deep learning model training.

Voyage

Participated in the large-scale integration of the TellApart codebase into the twitter ads ecosystem through the Website Conversions project which bridged advertising from twitter to the rest of the internet through display ads. Eventually joined the Ads Marketplace team which optimized the entire multi-billion dollar revenue ads products via the internal ads auction system.* Worked with twitter ML teams to experiment with deep learning approaches to improve the performance of our ads engagement prediction models. Implemented a BLAS-based inference engine for fast inference of plain multi-layer perceptron networks. Ran A/B testing to pit new deep learning approaches to previous models.* Invented a generic, nearly-optimal, binary encoding for ML feature vectors. Built an encoder/decoder library to integrate with spark to optimize batch job runtime and memory usage. * Instrumented the roll-out of the joint TellApart-Twitter ad product and built monitoring tools to oversee the performance across various dimensions as the product became more widely used.* Helped build and roll-out ad marketplace features to globally optimize revenue by matching bid pricing with expected traffic patterns.* Built a system to progressively filter out low-value advertisements based on a real-time feedback mechanism. This filter had substantial computation savings effects with a net-positive revenue impact after measured with A/B tests.

Twitter

- Acquired by Twitter)Worked on core RTB (real-time bidding) service, ML platform, ML feature-extraction/modeling, data analysis, and ad recommendations engine. * As part of the migration of the core bidding services to java, I overhauled the codebase into an async-friendly system using Futures, allowing for much greater parallelism across I/O bound requests as well as CPU utilization and latency reduction.* Factored out the ad recommendations system from a business-logic-laden legacy codebase into a well-instrumented, configurable microservice. This allowed for much greater debuggability, configuration-tuning, A/B experimentation, implementation of new recommendation engines, etc.* Built a ML features metadata service. The purpose was to conduct nightly feature extraction pipelines, describe available features, value histograms for regression monitoring, etc. Alongside this was various data consumption tooling to easily import datasets into ad-hoc or scheduled spark or hadoop batch jobs.

TellApart

Software Engineer

- Acquired by TellApart)Designed and built a real-time, ML-based decision engine to infer offer personalized experiences on online retail sites. * Built out the batch processing pipelines to extract features, retrain, validate, and deploy ML models.* Built library to share ML feature extraction logic in multiple contexts (batch processing vs online inference settings).* Factored out ML inference component into its own service. This greatly reduced hosting cost as we could use fewer, smaller frontend instances to handle incoming API requests, having a smaller number of more powerful machines running ML inference.

Freshplum, Inc.

Contributed to geometric tomography research project under Dr. Richard Gardner.