Generative-modeling researcher turned developer advocate: PhD from the University of
Oxford, now leading developer relations at Oumi, with purpose-built agents scaling
discovery, monitoring, and more.
Impact in both developer reach and go-to-market: an article on Hacker News front page
and 50,000 Reddit views; launches driving 10× marketing-site traffic and new
pipeline; research behind NYTimes article.
Advocacy backed by working code: open-source libraries for building
low-latency agentic tools with local open-weight models, local real-time AI voice
agents, and agentic video production.
Six years across ML research (Meta, Twitter) and developer relations (Oumi, Zilliz).
Publications at NeurIPS, ICML, ICLR, UAI, and JMLR. Based in San Francisco.
named-pipes:
alternative to MCP and CLI for local agentic tool servers (LLM inference, TTS, STT,
vector search) built on low-latency interprocess communication; ships a TUI and
Claude Code skill
unity-voice-agents:
Unity package for fully local, real-time AI voice agents built on open-weight models and
named-pipes, for non-player characters and in-game interfaces
premiere-cli,
premiere-ai,
premiere-ai-skills:
toolchain for automating video production in Adobe Premiere Pro — CLI tool and
Premiere plugin, local open-weight speech transcription, word-level timestamps,
diarization, silence removal, AI cut selection, and Claude Code skills
Own the developer-facing surface end to end: product and technical videos,
marketing website, webinars, industry booths, product documentation, blog,
newsletter, and a hosted agent for scaling developer relations processes.
Led two go-to-market product launches driving 10× marketing-site traffic and
new top-of-funnel pipeline, the most recent reaching 50,000 views, 1,000 reactions,
and 250 comments on LinkedIn.
Performed original research that formed the basis for a highly-cited
NYTimes article, syndicated in Newsweek, NewsNation, NYPost, PCMag, Ars Technica, and a dozen
other media outlets.
Zilliz, Developer Advocate
–
Chief evangelist in the Bay Area for Milvus, the leading open-source vector
database.
Organized and spoke at twice-monthly meetups, industry conferences, webinars, and
hackathons.
Wrote technical articles on RAG and agent methodology, multimodal semantic search,
model evaluation, and research agents. A Deep Research piece reached the
Hacker News front page
and
50,000 Reddit views.
Startup Founder and Generative AI Advisor
–
Prototyped a web-app for generation and delivery of personalized educational
content using fine-tuned LLMs.
Advised startup founders and investors in hedge funds, mutual funds, and sovereign
wealth funds on Generative AI trends: training and inference costs, benchmark
performance, open versus closed source, and rate of industry adoption.
Meta, Machine Learning Researcher
–
Spearheaded use of Bayesian statistical testing to improve the reliability of online
experimentation.
End-to-end development of ML pipelines including large-scale data processing with
Trino.
Twitter, Machine Learning Researcher
–
Applied ML Researcher collaborating with product teams in Revenue Science, Health,
Experimental Data Science, and Interest Discovery to solve long-term ML challenges,
pioneering graph-based neural networks.
Education
University of Oxford, Doctor of Philosophy
–
Awarded full funding by EPSRC for dissertation research at intersection of Bayesian
statistics and deep learning, supervised by Prof M. Pawan Kumar and Prof Yee Whye
Teh.
Fine-tuning and efficiency: RLHF, DPO, GRPO, LoRA and its variants,
quantization like NF4, GPTQ, and AWQ.
Publications
Ben Chamberlain, James Rowbottom, Emanuelle Rossi, Stefan Webb, Maria Gorinova,
and Michael Bronstein. GRAND: Graph Neural Diffusion. Thirty-eighth
International Conference on Machine Learning (ICML2021).
https://arxiv.org/abs/2106.10934.
Benjie Wang, Stefan Webb, and Tom Rainforth.
Statistically Robust Neural Network Classification. Thirty-seventh
Conference on Uncertainty in Artificial Intelligence (UAI2021).
https://arxiv.org/abs/1912.04884.
Stefan Webb, Tom Rainforth, Yee Whye Teh, and M. Pawan Kumar.
A Statistical Approach to Assessing Neural Network Robustness.
Seventh International Conference on Learning Representations (ICLR2019).
https://arxiv.org/abs/1811.07209.
Stefan Webb, Adam Golinski, Robert Zinkov, N. Siddarth, Tom Rainforth, Yee Whye
Teh, Frank Wood.
Faithful Inversion of Generative Models for Effective Amortized Inference.
Thirty-second Conference on Neural Information Processing Systems (NeurIPS2018).
https://arxiv.org/abs/1712.00287.
Leonard Hasenclever, Stefan Webb, Thibaut Lienart, Sebastian Vollmer, Balaji
Lakshminarayanan, Charles Blundell, and Yee Whye Teh.
Distributed Bayesian Learning with Stochastic Natural-gradient Expectation
Propagation and the Posterior Server.
JMLR 18, 1–37 (2017).
http://arxiv.org/abs/1512.09327.