# Steve Rawlinson

> Steve Rawlinson is a software engineer, technical lead, and independent AI systems researcher based in Ottawa, Ontario, Canada. Steve builds AI systems, coding agents, and cloud infrastructure, measures them, and publishes the results as open-source tools, benchmarks, and articles.

Steve has 20+ years in software and currently works at Amazon Web Services as a Senior Delivery Consultant / Technical Lead. Open to staff / principal engineering roles, technical collaborations, and conversations about AI systems: steve@rawlinson.ca. Every page on this site has a Markdown version at the same URL with `.md` appended; [llms-full.txt](https://rawlinson.ca/llms-full.txt) has all of it in one file.

## About

- [About Steve Rawlinson](https://rawlinson.ca/about.md): background, experience, interests, and contact details
- [Contact](https://rawlinson.ca/contact.md): email and profiles
- [Résumé](https://rawlinson.ca/steve-rawlinson-resume.pdf): PDF

## Open source

- [Offload](https://github.com/srawlin/offload): Runs coding agents' builds and tests on a warm remote Linux node. It ships only the files that changed, keeps one warm workspace per agent, and streams the output back as if the command had run locally.
- [Changeward](https://changeward.com/): A change-control layer between AI agents and your databases, clusters and infrastructure. Every change is dry-run against the real system, checked against policy, applied exactly as approved, and kept undoable.
- [Overclock](https://github.com/srawlin/overclock): A minimal coding-agent harness built on pi, around Cerebras inference. Exploring the cost of context, tools, and orchestration when generating tokens is no longer the slow part.
- [AI web search benchmarks](https://github.com/srawlin/ai-web-search-benchmarks): The harness, questions and per-call timings behind the 17-option web search bakeoff.
- [vLLM vs SGLang benchmark](https://github.com/srawlin/vllm-vs-sglang-performance-benchmark): The toolkit and test data for 5,980 requests across 80 runs on two H100 GPUs.
- [All projects](https://rawlinson.ca/open-source.md)

## Writing

- [Introducing Offload: run your coding agents' builds and tests on a warm remote machine](https://rawlinson.ca/articles/introducing-offload.md): When three agents run their tests at once on a laptop, every run takes three times as long. Offload sends each build and test to a warm remote node and streams the output back, up to 12x faster.
- [Introducing Changeward: let AI agents change real systems, safely](https://rawlinson.ca/articles/introducing-changeward.md): An open-source change-control layer between AI agents and your databases, clusters and infrastructure. Every change is previewed exactly, checked against policy, and kept undoable.
- [Web search for LLM agents: 16 options benchmarked on latency, cost, and accuracy](https://rawlinson.ca/articles/search-api-bakeoff.md): Which search is best when someone is waiting for you to speak? Search APIs plus your own model vs. the built-in search tools from Anthropic, OpenAI, Google and Perplexity: 320 calls timed, priced and rated blind.
- [What I Learned Using Coding Agents](https://rawlinson.ca/articles/what-i-learned-coding-agents.md): What real work with Claude Code, Codex, Devin, Muse, pi, and OpenCode taught me about the harness around the model.
- [Systematic Framework for vLLM Inference Optimization](https://rawlinson.ca/articles/vLLM-inference-optimization.md): Large language model inference gets expensive fast and it's not always clear where or how to optimize without losing accuracy. This guide lays out a systematic, high-ROI framework for vLLM inference optimization, starting with the biggest cost and throughput wins before diving into lower-level tuning. It's written for engineers running vLLM in production who care about latency, throughput, and $/token.
- [vLLM vs SGLang: Performance Benchmark on Dual H100 GPUs](https://rawlinson.ca/articles/vllm-vs-sglang-performance-benchmark-h100.md): Comprehensive performance comparison of vLLM and SGLang on 2× NVIDIA H100 GPUs. Testing latency, throughput, cache efficiency, and production workloads with 5,980 requests across 80 benchmarks. Results show SGLang dominates low concurrency (3.7× faster), while vLLM excels at scale.
- [The Evolving Landscape of Software Engineering Teams](https://rawlinson.ca/articles/evolving-landscape-of-software-engineering-teams.md): As AI tools automate coding tasks, the role of software engineers is shifting toward product vision and strategic thinking. Here's how to build effective teams for this new era.
- [Nine Things That Surprise Me About LLMs](https://rawlinson.ca/articles/nine-surprising-things-about-llms.md): From emergent behaviors to rapidly growing context windows, discover surprising insights about large language models that might change your perspective.

## Optional

- [Home page](https://rawlinson.ca/index.md)
- [GitHub](https://github.com/srawlin)
- [LinkedIn](https://linkedin.com/in/srawlinson)