Skylar Hoffman is a Staff Machine Learning Engineer at Moveworks (acquired by ServiceNow), where he leads enterprise search serving more than 200 customer environments — ranking systems, retrieval, and large language models in production. He holds a Master's in Computer Science (Artificial Intelligence) from Cornell University and a B.S. in Computer Science (Intelligent Systems) from UC Irvine. His research spans automated entity extraction from state regulations at Cornell and the detection of spiral galaxy arm segments at UC Irvine.
His working materials are Python, Go, C++, and SQL, applied to transformers, semantic retrieval, and distributed systems. Away from the keyboard he reads novels, plays volleyball, snowboards, and paints watercolor.
02
Skills
Tools and technologies.
03
Experience
Professional work, 2021 — present.
Staff Machine Learning Engineer
Moveworks — Mountain View · acquired by ServiceNow, 2025
2022 — Present
Led development and scaling of enterprise search across millions of resources in 200+ customer environments.
Architected multilayered ranking pipelines and advanced ranking models, improving NDCG by ~30%.
Integrated fine-tuned LLMs into production search and agentic workflows — query understanding, retrieval, summarization, and task completion — supporting an enterprise agentic AI rollout driving millions in net ARR.
Built evaluation and experimentation frameworks improving Cohen's Kappa by ~50% for automated LLM data labeling.
Mentored engineers and shaped technical direction for search relevance, retrieval, and LLM integration.
Software Engineer
MathWorks — San Francisco
June — August 2021
Built features for MATLAB application components used by a large global customer base.
Research
Cornell University · UC Irvine
2018 — 2022
Cornell (Prof. Frug, 2021–2022): machine learning models for automated entity extraction from state regulations corpora.
UC Irvine ICS Honors (Prof. Hayes, 2018–2021): SpArcFiRe — automated detection of spiral galaxy arm segments across multiple wavebands.
04
Expertise
Recent work at Moveworks, 2022 — present.
Enterprise Search
Production retrieval · 200+ enterprise environments
Inside a large company, the answer to a simple question rarely lives in one place — it is scattered across wikis, ticket histories, policy documents, and file stores, each behind its own permission walls. This work composes a single retrieval layer over that fragmentation: hybrid lexical and semantic search with permission-aware indexing, so that millions of documents across hundreds of enterprises surface instantly, and only to the eyes entitled to see them.
Proprietary work · Moveworks
Agentic Workflows
Fine-tuned LLMs · agentic AI in production
The Moveworks assistant is built not merely to answer but to act — provisioning access, routing approvals, and resolving IT and HR requests end-to-end through multi-step reasoning over plugins into hundreds of enterprise systems. Within that loop, Skylar's fine-tuned language models carry the conversation: interpreting what an employee actually needs, distilling what the systems return, and seeing the request through to its close.
Enterprise ranking affords none of the web's luxuries — no planet-scale click signal, and every customer arrives with its own corpus, acronyms, and internal dialect. These learning-to-rank models are trained on behavioral and structural signals drawn from each environment's own usage, yielding rankers that adapt to corpora they have never seen and lifting relevance, by NDCG, by roughly a third.
Proprietary work · Moveworks
Evaluation & Experimentation
LLM-as-judge · experimentation frameworks
How does one measure search quality across hundreds of private corpora, where no ground truth exists and no annotator may read the documents? The answer here is rigor at one remove — hand-built golden sets, LLM judges calibrated against human labels until their agreement, by Cohen's Kappa, rose by half — and an experimental discipline that demands evidence before any model reaches production.
Proprietary work · Moveworks
05
Projects
Independent and student works, 2018 — 2021.
Moogle Maps
A Minecraft agent trained with reinforcement learning to cross complex terrain.
