# Bryan Schwab

> Personal site of Bryan Schwab. Founder and engineer building with agents, with reports from the field.

Bryan is co-founder of Quirence AI and writes production code daily on its two products for teams building with AI agents: Rinkata (https://rinkata.dev), one source of truth for the goals, specs, and decisions a team and its agents share (launched on Product Hunt on September 30, 2026), and Pressure (https://pressure.dev), where agents from different labs debate a question before you commit. 25+ years across infrastructure, security, and payments. Previously Senior Director at Venmo (200+ engineers, $300B+ annual processing), Head of Security Software at Mysten Labs, and Engineering Manager at Meta (blockchain, security, infra).

This site collects writing and a post-race recap from the Tokyo Marathon 2026. Work history lives on LinkedIn.

## Pages

- [Homepage](https://www.schwab.sh/): Editorial cover with the latest post.
- [About](https://www.schwab.sh/about/): Bio, current work, and a link to the Tokyo recap.
- [Writing](https://www.schwab.sh/blog/): Blog index of essays on engineering, AI, and adjacent topics.
- [Tokyo 2026 recap](https://www.schwab.sh/tokyo26/): Post-race recap from the Tokyo Marathon.

## Writing

- [Which qwen Actually Fits a 16GB GPU](https://www.schwab.sh/blog/qwen-16gb-blackwell/): Benchmarking qwen on a 16GB RTX 5060 Ti; no qwen3.8 tag fits, two Modelfile lines make it 79% faster than the default, and qwen3.5 9B with MTP runs Q8 weights at Q4 speed.
- [The Best gemma4 12B Quant for a 16GB Blackwell GPU](https://www.schwab.sh/blog/gemma4-12b-16gb-blackwell/): Benchmarking gemma4 12B quants on the cheapest 16GB Blackwell card; QAT beats Q4, and the FP4, FP8, and MTP speedups are all macOS-only on Ollama.
- [I Let My AI Agent Plan My Tokyo Marathon Trip](https://www.schwab.sh/blog/openclaw-agent-tokyo/): An OpenClaw agent with its own Gmail, iCloud, and browser plans eleven days in Tokyo and Hakuba.
- [Benchmarking the Raspberry Pi AI HAT+ 2 for Local LLMs](https://www.schwab.sh/blog/hailo-ai-hat-benchmarks/): Testing the Hailo-10H NPU for home automation; 6–8 tokens/second on local models.
- [Claude Code Browser Automation on Bazzite](https://www.schwab.sh/blog/claude-code-browser-bazzite/): Setting up Chrome with Claude Code browser automation on Bazzite via distrobox.
- [Hello World](https://www.schwab.sh/blog/hello-world/): First post.

## External

- [GitHub](https://github.com/bschwab)
- [LinkedIn](https://www.linkedin.com/in/bschwab/)
- [X](https://x.com/bryanschwab)
- [Quirence AI](https://quirence.ai)
- [Rinkata](https://rinkata.dev)
- [Rinkata on Product Hunt](https://www.producthunt.com/products/rinkata)
- [Pressure](https://pressure.dev)

## Contact

- bryan@schwab.sh