[ Services ]
Senior engineer. Occasional advisory work.
I occasionally help small teams reason through backend reliability, Go performance, distributed systems, and production AI workflow design.
If the scope is technical, specific, and compatible with my availability, feel free to get in touch.
[ Areas_I_Can_Help_With ]
- Go backend performance and runtime-level debugging
- High-throughput service architecture: gRPC, NATS, Redis, queues
- AI agent orchestration, tool-calling reliability, and completion tracking
- Observability, latency, retries, deduplication, and cost control
- System internals perspective: runtime, memory, concurrency, and Linux-level diagnostics
- Architecture review for small engineering teams
[ Useful_When ]
- Your Go backend slows down under load, but pprof, logs, and dashboards do not explain why.
- Your AI agent workflow works in demos, but loses state, retries incorrectly, or costs too much in production.
- Your team is deciding between RPC, queues, streams, or event-driven architecture and wants a second technical read.
- Your system has reliability issues around latency, retries, deduplication, observability, or cloud cost.
- You need a senior engineer to review architecture before scaling a small engineering team.
[ Active_Engagement ]
AI Agent Production Readiness Review
A focused review for teams shipping AI agents into production.
Two to three weeks of work. I walk through the seven failure modes that account for most production AI agent issues — open-loop validation, unbounded context growth, retry storms, tool-call completion ownership, state management drift, latency budget enforcement, cost attribution — against your specific codebase, and produce a prioritized fix list with concrete code references and expected impact.
What's included
- Walkthrough of the seven failure modes against your specific agent codebase
- Prioritized fix list — each item with code reference, expected cost / latency / reliability impact, and effort estimate
- Architecture review covering validation loops, tool-call ownership, state stores, retry budgets, observability
- Written report you can hand to your team for implementation
Typical signals before reaching out
- Your agent reports success more often than user-visible outcomes confirm
- Monthly LLM bill is 3–5× what your per-call token meter predicts
- The same input produces different outputs across "identical" sessions
- You've shipped a working demo but reliability degrades at customer scale
For scoping: [email protected] — a short note about what you're working on is the easiest way to start.
[ Recent_Writing ]
A few representative pieces.
Claude Code Source Leaked: 5 Hidden Features Found in 510K Lines of Code
What Anthropic's leaked agent source reveals about how production AI agents actually work.
Claude Code · ArchitectureThe 1,421-Line While Loop That Runs Everything
10 steps per iteration, 9 continue points, 4-stage compression, streaming tool execution.
Claude Code · Context EngineeringThe 5-Level Compression Pipeline Behind 1M Tokens
How Claude Code keeps sessions alive past the context window without losing fidelity.
AI Agents · ReliabilityThe 90% Problem: Why Most AI Agents Are Still Broken
The reliability gap between “works in the demo” and “works on Tuesday at 3am.”
AI Infra · ObservabilityObservability and Billing for AI API Calls: A T-Shaped Architecture
Why your AI bill is 3-5x what your token meter says, and the architecture that fixes it.
Go Runtime · ConcurrencyNode Turns Waiting Into Events. Go Moves Context Switching Into User Space.
Stackless vs stackful coroutines, function color, and what the unit of scheduling actually is.
Distributed Systems · PatternsRPC vs NATS: It's Not About Sync vs Async — It's About Who Owns Completion
The framing that separates teams who debug at 3am from teams who don't.
Go · ProfilingGo Profiling in Anger: pprof, Escape Analysis, and Inlining Without Magic
The exact tools and reasoning chain I bring to Go performance work.
[ Get_in_Touch ]
A short note about what you're working on is the easiest way to start.