Luis OtavioLuisOtavio.

Problem

Tapnote, a Polish AI study app with 30,000+ subscribers, lived on mobile. Students needed the same tutor on the web.

What I built

• Built the web app on my own, frontend to backend, at feature parity with the native mobile app.
• Built AI tutoring with LLMs and RAG on PostgreSQL / Supabase for homework help and personalized study assistance.

Problem

Enterprises such as Siemens Energy, Henkel and Essity needed a training platform where courses could be drafted automatically, with each organization's users and data strictly isolated.

What I built

• Built LLM course generation and AI-assisted authoring tools, so a course starts as a draft instead of a blank page.
• Owned multi-organization authentication and tenant isolation with Better Auth.
• Designed and shipped a peer review system so training creators receive structured feedback and ratings.
• Implemented ISO/IEC 27001 controls with Kertos, supporting the company's certification.

Problem

Volunteer fire brigades in central Brazil and Brazil-nut collectors in the north both coordinate in WhatsApp groups, far from the platforms where their data lives: Caminho do Fogo for fires, IE Sociobio for harvest prices. I build the AI agents; Terra Krya brings them to the communities.

What I built

• Curupira: brigade members sign in from the chat, report alerts and fires with a location pin and photos (each write confirmed before it is saved), check the INPE fire-risk forecast and ask questions answered by RAG over 68 brigade documents, with citations.
• OCA IÊ: producers log harvest prices just by talking and compare regions; a weekly regional recap goes to each group, rewritten or rescheduled by asking the bot.
• Both forked from my open-source AI Agent Boilerplate: one Python agent with voice notes in and out, on WhatsApp (Curupira on Telegram too), 2,300+ pytest and Vitest tests and Sentry monitoring.
• FleetView, a Next.js dashboard I built to watch the fleet's health, conversations and knowledge bases.

Problem

Travelers either book a fixed group tour or piece a day together from blogs and maps. TourLore is an AI tour guide and trip planner for iOS and Android: it writes a personal walking tour for any city, narrates each stop, works offline and answers photo questions mid-walk.

What I built

• Expo / React Native app with Expo Router, NativeWind, MapLibre maps, audio narration and offline mode, in a Turborepo monorepo with EAS builds.
• Backend: a Hono + tRPC API on Google Cloud Run, Zod contracts shared with the app, Drizzle on Neon PostgreSQL, Inngest background jobs, Sentry and GitHub Actions CI.
• An LLM tour-planning agent on the Vercel AI SDK and Gemini on Vertex AI, with tool calling, quality-gated by LLM-as-judge evals in Langfuse.
• The Next.js 16 landing site in 30 languages (next-intl), with a Resend waitlist.

Problem

When a change spans repos, you either restart in each repo and explain everything again, or let one AI agent edit a repo whose rules, hooks and conventions it never loaded.

What I built

• A Python CLI where a lead agent hands each repo's part of a change to a Claude Code or Antigravity agent working inside that repo, in its own git worktree, with that repo's rules loaded.
• Plan, review, execute: tasks are sized, strong models plan, the lead reviews each plan before any edit, and the same conversation executes on a cheaper model, switching models when a quota runs out.
• Every PR gets cross-agent reviews from Claude Code, Gemini (Antigravity) and GitHub Copilot, and every agent runs in a terminal pane you can watch and type into.