senior sre, canada-based. 12 years keeping distributed systems alive — kubernetes, terraform, monitoring, incident response, data platforms. that background turned out to be the useful part of ai: models are easy to demo and hard to operate, and operating things is what i already do.
lately that means making llm systems production-grade:
aissist — incident summarizer running real customer text through azure openai. rag over incident history cut resolution from 30 days to 10. a tiered pii guard sits at the chokepoint, and the secure-llm-use pattern it proved replaced per-team security reviews.
hider — redaction gateway for log bundles on their way to a model. presidio patterns, a local model catching the residue, and a proxy that refuses to forward what it could not clean.
llm-batcher — an openai-shaped proxy in front of claude. translates both ways, microbatches concurrent requests, and measures its own cost and latency.
llm-intelligence-layer — a multi-agent runtime where advisory traffic is read-only and acting traffic passes through a human approval that survives restarts. plan → approve → tools → publish.
arbiter — fuses findings from two security vendors, ranks them by blast radius across the estate, and lets no suppression through without a deterministic check the model cannot argue with.
around those: groundloop for llm service-account management on macos, imaginator for provider-pluggable image generation, explanator and tg-fetcher for models woven into everyday tools.
the pattern in all of it: evals before rollout, observability in the request path, cost under control, and a guard that doesn't trust the model to behave. that's usually what's missing when pilots don't ship.
master's in computer science. applied software development certificate with distinction. python, go, swift, terraform, kubernetes, linux, cloud.
this site is a directory of the tools and write-ups i ship on the side. nothing is for sale. most of it is open.
if something here is useful, or wrong, or both — i'd like to hear about it.