I started at Oracle’s Bronto division writing MapReduce jobs against billions of records and cleaning up data systems that had outgrown their original design. At Bandwidth I led a team across reporting, deployment pipelines, and a geo-redundancy project that had me writing Ansible playbooks, coordinating datacenter failovers, and doing postmortems before company-wide outage review boards.
At Toast I work on the Data Platform team — the analytics infrastructure that processes billions of restaurant transactions. Most of the company depends on it; nobody thinks about it until it breaks. My job is to make sure it doesn't, and to build the tooling that makes diagnosing it fast when it does.
Analytics pipelines that process restaurant transaction data at scale. Iceberg table management, query materialization, and infrastructure across multi-account AWS environments.
MCP servers for persistent memory, investigation agents for automated debugging, orchestration frameworks, and skill-based architectures for Claude Code.
CLI utilities, automated investigation systems, deployment automation, and observability tooling that bridges the gap between systems and the people who operate them.
Google Research’s three-paper family — QJL, PolarQuant, and TurboQuant — compresses LLM KV caches 6× with zero quality loss using one elegant trick: random rotation eliminates quantization overhead entirely. vLLM and SGLang had working PRs within four days.
Karpathy’s autoresearch loop, harness engineering as cybernetics, Codex 5.4 reverse-engineering a DOS game in 6 hours, and an agent called Larry running a $7k/month business autonomously. Curated for builders who ship.
Move agent behavioral control out of CLAUDE.md and into deterministic code. Four composable patterns — PreToolUse interception, orchestrators, session state, and PostToolUse context injection — that fire reliably regardless of context window pressure.
I built a Wingspan board game simulator with 446 birds and a hand-tuned AI full of hardcoded thresholds. Then I read a paper where Google let an AI rewrite game-theory algorithms — and it found better ones. Here’s what that would look like applied to my project.
Google DeepMind pointed an AI code-evolution agent at game theory — the math behind poker bots, autonomous negotiation, and multi-agent AI. It discovered two algorithms that outperform 20 years of hand-designed ones. A visual explainer of what it found and why it matters.
Every AI chatbot re-reads your entire conversation to generate each word. DeepSeek found a way to skip the parts that don’t matter — three reading strategies (summarize, select, remember) that cut processing time by up to 11× without losing quality. A visual explainer of the paper.
Zhipu AI’s 744B parameter model competed anonymously as “Pony Alpha” on OpenRouter — users thought it was Claude Sonnet 5 or DeepSeek. A visual explainer covering the architecture (Dynamic Sparse Attention), training pipeline, benchmarks, and why it matters for open-source AI.
Claude Code is built for Unix-like systems. On Windows, you need WSL2 — without it, Claude Code can’t run scripts and resorts to generating .bat files. A step-by-step guide from install to full autonomous execution.
GitNexus indexes a codebase into a knowledge graph — 13k+ nodes and 38k+ edges. Query execution flows, check blast radius before refactoring, and let AI agents see graph context automatically on every search.
A survey of the AI-assisted development landscape. Seven in-depth profiles covering orchestration tools from Claude Squad and Gas Town to Conductor and TSK — how each approaches multi-agent coordination, context management, and developer workflow integration.
Kubernetes for AI agents. Mayor + 20 workers in tmux with beads for persistence.
→Brownian ratchet — always pushing forward. Supervisor + subagents with auto-merge.
→Plan, Work, Assess, Compound. Open-source plugin and most-referenced methodology.
→Terminal app managing multiple Claude Code instances in separate workspaces.
→Native Mac app. Visual dashboard with git worktree isolation and rollback.
→Rust CLI delegating tasks to agents in sandboxed Docker containers.
→Docker-based dev environments for coding agents. Isolated parallel execution.
→The complete landscape — who's using what and how each stack compares.
→