I work where AI systems meet real codebases — taking something that works in a demo and making it hold up in production, inside the systems people already run.

That usually means working inside other people's projects, not just my own: reading the existing code, reproducing the problem, writing a small fix that fits the project's conventions, and iterating through review until a maintainer is comfortable shipping it.

A deterministic, offline CLI that audits LLM projects for production readiness across five areas — evaluation, guardrails, observability, governance, and reliability. 15 checks, scored GO / CONDITIONAL GO / NO-GO reports, with a passing and a failing sample included.

• https://github.com/trleo/llm-production-readiness-audit

A CLI that audits, deduplicates, and prunes AI agent skills across Claude, Cursor, and Antigravity. Duplicate detection by name, content hash, and similarity — with approval-file-only pruning, quarantine with a manifest, and restore. It never deletes.

• https://github.com/trleo/skill-cleaner

• chunkhound — SSRF protection in websearch (#475): added a shared SSRF validator so search-result URLs are checked before fetching, including hop-by-hop redirect validation, secure-by-default settings, and behavioural tests. Approved after three review rounds and merged into main.
• scalar — Auth badge overflow (#10841): fixed a floating panel that briefly overflowed the viewport while hovering an "Auth Required" badge, with a regression test. Merged the same day it was opened.
• Ovenbird — cos power command (#133): a new read-only command reporting macOS power assertions. Ovenbird's first outside contribution; thanked in the 0.8.0 release notes.
• Ovenbird — Vitals trend sparklines (#157): 30-day trend sparklines for Vitals, with a collecting state while data accumulates.
• rollup — Exported destructuring with defaults (#6648): fixed invalid output for exported variables in destructuring patterns, with a live getter for reassigned variables. Merged into master.

Open pull requests under review with maintainers, including fixes for good_job (NULL queue/job-class charts), rclone (archive backend move-onto-itself), tenacity (LoggerProtocol compatibility with loguru), vite (CSS HMR query handling), glow (style-path expansion), vLLM (MFU/MBU for renamed quantisation methods), TheAlgorithms/Python, and skills-janitor.

• Reproduce first, then fix — in the project's own test setup, not in isolation.
• Small, reviewable changes that fit the codebase's existing patterns.
• Verified end to end: tests green, honest about what could not be reproduced locally.
• Security and reliability fixes first — the kind of work that has to be right before it ships.
• Python first, with working contributions across Go, TypeScript, and Ruby projects.

• WhatsApp: +8801737923848
• LinkedIn: linkedin.com/in/trleoo