Skip to content
View felipefrf's full-sized avatar

Block or report felipefrf

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
felipefrf/README.md

felipefrf.com: I build AI agents that do real work. A live orbital simulation of the satellite my ITA team designed.

Hi, I'm Felipe 👋

AI engineer from ITA. Founding engineer at Hexagon, where I build Milo, an AI teammate you hand work to.

felipefrf.com: my site, with a live orbit simulation, the systems I build animated end to end, and the work below in more detail · LinkedIn · São Paulo, Brazil


Now: Milo

Every person on the team gets their own agent, running in its own Linux microVM with their apps and logins. You give it the job and it carries it through, instead of just answering. My parts of it:

  • Voice. Live calls with Milo, on the phone and in the app. It keeps working while you talk, and you can cut in and steer it.
  • Where people already work. Microsoft Teams and Google Chat on par with Slack and WhatsApp: act-as-user OAuth, streaming, files and voice notes.
  • Credentials. An encrypted vault for the hosted browser agent. Passwords and one-time codes go in by alias and never reach the model.
  • Evals before shipping. Bilingual ship gates that replay real cases, nightly LLM-judged reviews, and an "autopsy" tool that rebuilds any conversation end to end.

The company code is private. Happy to walk through the architecture on a call.


Things I've built

The chief of staff's approval queue on a phone I built my own chief of staff. It reads WhatsApp, Slack, e-mail, meetings and GitHub, works out what needs me, and has the answer, the e-mail or a whole working session ready. I only approve, from my phone.
  • Reads everything: WhatsApp straight from the Mac's database (read-only, allow-listed contacts), Slack, three Gmail accounts, Granola after each meeting, calendar, Drive and GitHub.
  • Decides what needs me: a Claude session on my server loops every 15 minutes and wakes the moment I message it. It remembers people, projects and how I write, and learns from every draft I reject.
  • Prepares the work: replies with the reason behind them, e-mails, the daily's agenda. Anything bigger becomes a lane, a background Claude session with a brief that opens the PR or produces the PDF.
  • I approve, it ships: one queue on my phone with push notifications. Nothing goes out without me.
7,600+ Claude Code sessions and 22,000+ subagents in the last 30 days, across 7 accounts in parallel.

Claude Code MCP Node SQLite Web Push Cloudflare Tunnel

See the architecture →
penpad: the same sentence handwritten six times Markdown + LaTeX notes turned into realistic handwritten pages for iPad, Goodnotes and reMarkable. Fractions, integrals and vectors in handwriting. Offline, pip install, no API key.

Python resvg LaTeX
Curo home page A WhatsApp patient-care layer for specialist doctors: one chat for the patient, while the agent deals with the hospital, the insurer and the referring doctor. Built with my brother; got us a Y Combinator interview.

Google ADK Gemini Postgres WhatsApp

Smaller tools

  • editable-pets: a macOS companion that keeps every Claude Code and Codex session in sight.
  • yt-to-md: any YouTube video as clean Markdown for AI agents. No API key.
  • dual-subtitles-extension: two synced subtitle languages on YouTube, for learning a language.
  • x-reset: bulk-delete old tweets from your X archive. Resumable.

Stack

TypeScript, Python, React, Node.js, FastAPI, Postgres, GCP, AWS, Terraform, Docker

LLM agents and orchestration, voice, evals and guardrails, RAG, browser automation.


Before this

NG.CASH: production LLM agents for a Brazilian fintech, including a WhatsApp debt-negotiation agent that never lets the model pick the number. Credix: credit models that became the automated decision engine. LinkedIn: an LLM chatbot for the US & Canada sales forecast.

Before code, physics: 1st overall at the Iberoamerican Physics Olympiad 2020, 50+ olympiad medals.

Pinned Loading

  1. Livus-AI/zap-web Livus-AI/zap-web Public

    Zap website

    TypeScript

  2. dual-subtitles-extension dual-subtitles-extension Public

    Chrome (MV3) extension that shows two synced subtitle languages on YouTube — a reference and a learning language. Built with WXT, React, and a Shadow-DOM overlay.

    TypeScript

  3. software-house software-house Public

    Institutional website + automated lead-prospection system for a software house (Ápice).

    TypeScript

  4. x-reset x-reset Public

    Bulk-delete your old tweets, retweets, and replies from X/Twitter from your official archive. Playwright-driven and resumable.

    TypeScript

  5. yt-to-md yt-to-md Public

    Turn any YouTube video into clean, structured Markdown — built for AI agents. No API key, no cost.

    Python