i build backend systems and ai agents — the kind of software where the hard part is invisible and the bugs only appear after midnight. conveniently, that's when i start working.

off the clock i'm into paranormal investigation, which turns out to be the same job as debugging: something impossible happened, nobody witnessed it, and the logs are useless. i believe every haunting has a root cause. so far, it has always been a race condition.

also: cats. 🐈‍⬛

other people's codebases, other people's hauntings. 7 PRs merged across 6 upstream projects, three of them cncf.

• apicurio registry — #8684 merged. then found #8926: MCP tool annotations were being validated against the wrong schema entirely, so not one of the four hints anyone wrote was ever checked. fix in review — #9055.
• opentelemetry _{go compile-time instrumentation} — #856 merged. then found #886: struct rules cheerfully matched things that were not structs, and inside a grouped type block a field could land on a struct nobody named. fix in review — #887.
• kubeflow pipelines — #13780 merged: the python sdk was quietly adding --trusted-host to every pip call. opt-in now.
• microsoft/vscode — #324571 and #324734 merged in v1.129.0: a rerun-last-task regression, and a pinned tab that refused to let go when dragged.
• nodejs/node — #64365  ·  fedora-infra/bodhi — #6121
• _{in review: kserve #5908 · #5909  ·  textualize/textual #6631 · #6630}

the half i like best: find it myself, report it, then fix it.

a competitive programming platform, built end to end — 95+ problems, an async code judge that speaks 10 languages, real-time verdicts over websockets, elo ratings, plagiarism detection.

the favourite child: a custom segment tree that can replay the leaderboard at any minute of any past contest. time travel, but auditable.

_{typescript · nestjs · next.js · postgresql · redis · bullmq  —  source}

security incident response, run by two LLM agents keeping each other honest — a Commander proposes actions, an Oversight agent approves, revises, or vetoes.

both agents trained with GRPO reinforcement learning on qwen2.5-3b. placed top 100 at the meta pytorch × scaler openenv hackathon.

_{python · pytorch · grpo · docker  —  source}

search-as-you-type, built like it has to survive production — a redis cache sharded across 3 nodes on a consistent-hash ring, an in-memory trie serving top-K suggestions, a batched writer that collapses thousands of db writes into a handful.

benchmarked properly: p50, p95, p99.

_{python · fastapi · redis · sqlite}

a sales agent that haunts your google calendar, benevolently. thirty minutes before every meeting, it researches the prospect on the live web and emails you a brief with cited sources.

nine google cloud services stitched together, zero manual invocation. it just shows up, prepared — like you wish you did.

_{next.js · gemini · cloud run · firestore}

a poltergeist you can hire. it drives a real chromium browser with actual mouse and keyboard events, finds form fields through the accessibility tree, and streams every action and screenshot to a live dashboard.

so you can watch the ghost work.

_{python · playwright · fastapi · docker  —  source}

an internet speed test that doesn't try to sell you a VPN. download, upload, ping, jitter, bufferbloat — measured against cloudflare's edge with zero backend, rendered on a hand-rolled SVG speedometer.

_{react · vite · typescript · cloudflare pages  —  source}

• global top 6 — apple developer academy, italy
• top 100 — meta pytorch × scaler openenv hackathon

typescript python go java c++ nestjs fastapi quarkus express react next.js websockets postgresql redis mongodb sqlite pytorch grpo gemini · claude · openai docker gcp cloudflare vercel linux

 /\_/\
( o.o )   
 > ^ <