Dylan Ma

Simulation environments to train & evaluate long-horizon AI agents. We work with leading model labs to push the frontier of agent capabilities. Polymath is backed by Base10, Founders Future, Y Combinator, and other incredible investors & angels. We've raised an $8M seed round, and are actively growing out the founding team. We’re hiring engineers, researchers, and operators interested in advancing safe, highly capable AI agents: https://www.polymathlabs.ai/careers

- Training frontier voice models: EVI (conversational voice AI), OCTAVE (text-to-speech), voice creation (text description to voice) - SFT / RL / Data https://www.hume.ai/

Built and shipped the first voice AI agents into production

- Large scale distributed systems for ML model inference at AWS SageMaker. Launched multiple features into production to enable running generative AI workflows on AWS. - Inference response streaming: part of a 2 person engineering team to enable streaming inference responses from SageMaker endpoints. This allows pieces of the response to be returned to the client as soon as they become available, instead of waiting for the entire response to be generated. This is a critical feature for LLMs, and is used by AWS Bedrock as well as many external customers. - Load aware routing: led the API design for a feature that intelligently routes traffic to endpoint instances based off of a load heuristic. This has been shown to reduce e2e latency of LLM inference responses by more than 30%, and is currently being used in production by large customers. Received the Amazon Inventor Award for this work. - Multi-replica endpoints: worked on a feature that allows customers to deploy hundreds of foundation models on the same infrastructure. Models are optimally placed to maximize instance utilization, and scale independently according to demand. This leads to a significant reduction in deployment costs. - Distributed graph inference: worked on a feature that allows customers to deploy a computation graph of logical nodes, where each node is a cluster of instances that auto-scales according to demand. - Region build automation: launched the SageMaker Inference platform in 4 new AWS regions, ap-southeast-3, me-central-1, us-gov-east-1, and ca-west-1. Helped automate the process, reducing the dev effort for launching future regions by more than 90%. - Helped maintain a large-scale distributed system with millions of TPS and a 99.9% availability SLA. Addressed root causes for platform issues, improved internal tooling, and answered customer questions.

- (Now SKY Lab). Research in ML systems

- Research in AI & robotics

- Worked on AWS SageMaker Model Inference, a highly available, low latency service for hosting ML models and serving predictions. - Designed and implemented a new feature for AWS SageMaker Model Monitor that allows customers to detect outliers and distribution shift for computer vision models.

Highlighted coursework: Artificial Intelligence, Deep Learning, Operating Systems, Robotics, Efficient Algorithms and Intractable Problems, Back-End Web Architecture, Database Systems, Computer Architecture, Data Structures, Discrete Math and Probability Theory, Structure and Interpretation of Computer Programs, Designing Information Devices and Systems I & II, Biomedicine for Engineers

Admitted to Yale on an NSF (National Science Foundation) Fellowship. Decided to pursue research in industry instead