hi, i'm Thierry — a software engineer from São Paulo, Brazil 🇧🇷.

currently working at Meistrari, building Tela. i work across the stack: SDKs, document parsing pipelines, workflow design, and a workflow DSL for defining workflows as code while keeping them synced with the UI for developers and power users.

lately i've also been making codebases more ready for coding agents: better repo context, validation paths, harnesses, and practices that make agents useful instead of chaotic.

• building Tela
• tuning my AI developer setup around GPT-5.5 high/xhigh and pi
• building my own pi setup and extensions after a lot of time with Codex
• learning distributed systems by building them
• experimenting with microVMs, sandboxes, and local agent isolation
• using nvim btw

• microvm — local Firecracker microVMs from Dockerfiles for isolated coding-agent and dev workloads.
• dev — my reproducible dev setup: shell, editor, agent configs, and pi extensions under configs/agents/pi/extensions.

i'm building my own pi extensions, inspired by existing tools, because i believe developers should own this part of their workflow. i keep a strong focus on testability so the setup does not break while i'm using it.

• desktop-notification — an agent_end hook that emits OSC 777 desktop notifications when an agent turn finishes. see it
• web-access — web_search, fetch_content, and get_search_content tools for research-heavy agent work. see it
• create-image — /create-image for prompt-to-image generation inside pi. see it
• oracle — oracle for ChatGPT Web / GPT-5.5 Pro second opinions on hard problems. see it
• agent-feedback — agent_feedback for durable workflow feedback and verification blockers. see it
• subagent — a Claude-compatible Agent tool for single and parallel child Pi subagents. see it
• statusline — footer additions for clickable PR numbers, git change counts, and the Cloudflare BDR quote. see it
• token-speed — real-time assistant streaming throughput in Pi's footer. see it
• codex-fast-mode — injects Codex Fast mode's service_tier: "priority" for eligible ChatGPT-backed Codex requests. see it
• codex-compaction — uses Codex native opaque compaction for /compact on ChatGPT-backed Codex models. see it
• nested-agents — dynamically loads nested AGENTS.md / CLAUDE.md files as relevant files are touched. see it
• lsp — Language Server Protocol diagnostics and source-fix tools for Pi. see it
• fff — FFF-backed grep, find, and multi_grep search and autocomplete. see it
• excalidraw-session — an always-on local Excalidraw bridge tool for the focused browser canvas. see it
• goal — persistent autonomous goals with safe continuation, high turn limits, and mandatory completion auditing. see it

• coding agents and harness engineering
• distributed systems and systems programming
• microVMs, Firecracker, isolation, and sandboxes
• software engineering under the hood
• 3D printing