Svetlana Karslioglu
Information Solutions Engineer at Meta. Developer documentation, content systems, AI-native experiences.
10+ years building docs platforms, pipelines, and content for products engineers use directly. Currently at Meta on PyTorch, working on docs.pytorch.org and the AI-native experiences layered on top of it.
Where I've worked
Meta, Information Solutions Engineer (PyTorch)
- Supporting PyTorch releases. Launch planning, keeping docs.pytorch.org in sync with the API, and the Sphinx docs-as-code pipeline that lets engineers contribute directly
- Implementing AI-native solutions. Ensuring the docs are accessible and parsable by LLMs, building the retrieval layer that sits on top of them, and using Claude to scan, draft, and audit at corpus scale
- Co-authored The 99% Success Paradox: When Near-Perfect Retrieval Equals Random Selection (accepted to the ICLR 2026 blog post track), on how production RAG systems can report near-perfect retrieval while actually selecting documents at random
- Created and maintain pytorch-sphinx-theme2, the Sphinx theme used by docs.pytorch.org
- PyTorch Docs Maintainer
Pachyderm, Senior Technical Writer
Documentation for a Kubernetes-native data versioning and ML pipelines platform. Wrote tutorials, conceptual guides, and API reference. Published Reproducible Data Science with Pachyderm. Worked directly with engineering and product on launch content.
Rackspace, Information Developer
Documentation for cloud and container products, including Kubernetes-as-a-Service. Improved site search and UX on the docs platform. Shipped content every sprint on an agile engineering team.
Mirantis, Information Architect and Senior Technical Writer
Defined the information architecture and content standards for an OpenStack-based cloud platform. Helped design the open-source CI/CD docs toolchain (Sphinx, Gerrit, Jenkins). Upstream OpenStack contributor. Speaker at OpenStack Summit Austin 2016.
Selected work
The 99% Success Paradox: When Near-Perfect Retrieval Equals Random Selection
Introduces Bits-over-Random (BoR), a chance-corrected measure of retrieval selectivity. Shows that on 20 Newsgroups, BM25 and SPLADE both report >99% success at K=100 while actual selectivity is near zero, and downstream RAG accuracy degrades accordingly. The same collapse shows up in LLM agent tool selection when the catalog is small.
Reproducible Data Science with Pachyderm
Practitioner guide on versioned data pipelines for ML workflows.
Cisco Integrated Desktop Virtualization Solution
Solution architecture publication on enterprise VDI.
Speaker, OpenStack Summit Austin 2016
Tools and topics
- Python
- PyTorch
- Sphinx
- Docker
- Kubernetes
- Docs-as-code (CI/CD)
- LLM-powered content workflows
- RAG
- Information architecture
- Technical writing and editing
- Developer experience
Where I studied
University of California, Santa Cruz. Certificate in Technical and Scientific Communication.
Kuban State University. Master's Degree, International Relations.
Spoken languages
- English (Native or Bilingual)
- Russian (Native or Bilingual)