Peter Salas
CTO at Ultravox.ai
Seattle, Washington, United States
Summary
🤩
Rockstar🎓
Top SchoolPeter Salas is a Seattle-based technology leader and CTO with 12 years of experience building high-performance systems and machine learning infrastructure. He progressed from core browser and rendering work at Microsoft to platform and ML-focused roles at OctoML and Ultravox.ai, combining deep systems engineering with product-oriented delivery. At Ultravox he moved from software engineer to CTO, steering the company’s audio-first ML product strategy while contributing core engineering work. Peter contributes to popular open-source ML tooling—adding multimodal audio support to vllm, a high-throughput LLM inference engine—demonstrating hands-on expertise in integrating audio feature extraction and inference pipelines. He brings a rare mix of low-level performance optimization and modern ML deployment experience, grounded in a Computer Science degree from Harvard. Colleagues rely on him to translate research-grade models into production-ready, scalable services.
12 years of coding experience
15 years of employment as a software developer
A.B. Computer Science, A.B. Computer Science at Harvard University
Github Skills (12)
pytorch10
transformer10
audio-processing10
model-management10
inference10
multimodal10
python10
llm10
pytest9
testing9
deep-learning9
openai-api8
Programming languages (8)
Github contributions (5)
vllm-project/vllm
Aug 2024 - Feb 2025
A high-throughput and memory-efficient inference and serving engine for LLMs
Role in this project:
ML Engineer
ML EngineerContributions:58 reviews, 14 PRs, 73 comments in 5 months
Contributions summary:Peter primarily contributes to the development and testing of multi-modal features within the VLLM framework, specifically focusing on audio language models. They added support for Ultravox, an audio model, integrating it into the existing architecture, and implemented necessary input processing and mapping functions. Their work involves adapting existing code to incorporate audio data, including audio feature extraction, embedding, and integration with the model's inference pipeline. The user also addressed related issues like multi-modal placeholder tracking and handling of multiple audio chunks within prompts.
amdcudadeepseekgpthpu
fixie-ai/vllm
Aug 2024 - Nov 2024
A high-throughput and memory-efficient inference and serving engine for LLMs
Contributions:1 PR, 61 pushes, 16 branches in 2 months
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.