GabrielCamara Carvalho

• Coding assistants

  • Claude Code
  • Codex CLI
  • Cursor
  • Continue
  • Kimi Code
  • Manus AI

• Models and inference

  • Claude API
  • Qwen via Ollama
  • LM Studio
  • OpenAI-compatible local servers

• Agents and orchestration

  • Agent skills, hooks and subagents
  • MCP servers
  • Hermes Agent
  • Paperclip
  • Odysseus
  • Agent Town

• Memory, context and tooling

  • ai-memory
  • Graphify knowledge graphs
  • RTK
  • LanceDB RAG
  • Claude Design canvases

• Skills

  Skills turn the boring, error-prone parts of a workday into one command, and every step is verified against the real code. One reads a ticket and checks it against the codebase before any code is written, flagging inconsistencies. Others refine a ticket into a user story with acceptance criteria, run an implementation end to end, open PRs across several repos with clean commits, backport a change across release branches and write a 'how to test this change' guide. There is also a review-and-implement loop for my own mobile app, an evidence-based profile of a developer built from their review history, and a cleaner that strips AI tells from writing.

• Hooks

  Hooks keep the output honest without depending on my memory: RTK, a token-filtered shell that cuts 60 to 90 percent of what the model would otherwise read; a prompt classifier that forces a design interview before a new feature and a diagnosis before a fix, and offers to fan the work out to parallel agents; and automatic memory capture across the session lifecycle.

• Specialist agents and second opinions

  One specialist per repository: a mobile app, an admin panel, a coach app and a Kotlin/Flowable workflow engine, each with its own context. They run in parallel on disjoint files, so a change that crosses repos is split, built and verified at the same time. Codex CLI runs as an independent second pass on diagnoses and reviews, so two models have to agree before I act.

• Scheduled runs

  Every morning, before the workday starts, one run triages my ticket board and another reports on yesterday's work, so I open the day already knowing what is blocked and what moved.

• Memory and context

  A persistent memory with hundreds of notes on decisions, traps and verified findings, shared across agents through ai-memory; Graphify knowledge graphs of the codebases I work in; and design canvases used to settle direction before anything gets built.

• Local models and a self-hosted stack

  Local models were the starting point and never left: Qwen through Ollama and LM Studio on a Radeon iGPU and on a VPS, powering a RAG assistant that answers questions over a multi-repo codebase and a writing assistant that runs fully offline. Around them I run my own agent stack: Hermes Agent as the worker, Paperclip as the orchestrator, an Odysseus workspace, reachable behind Cloudflare Tunnel and Tailscale, plus a small MCP server I wrote that gives assistants direct Postgres access to my Supabase project.

• iOS · Android · Apple Watch · Web

  Health Tracker

  Train, recover and understand your own data, on the phone and on the wrist.

  A training and recovery app for phone and Apple Watch: structured workout execution, sleep, readiness and recovery guidance, correlations across sleep, training, mood and weight, strength progression, AI blood-test extraction, coaching and a social layer, with a premium subscription.

  TestFlightMobile app

  See the screens

• Web

  Agent Town

  Watch your AI agents work, and talk to them, in a 3D office.

  A Gather-like 3D office for watching AI agents work: running agent sessions are the characters, each live session gets a room, subagents walk in as named engineers, and you walk around with WASD, read the room log and type instructions to the agents. Ships an MCP server so agents can join, speak, move and check their inbox.

  Open sourceWeb app (3D)

  GitHubSee the screens

• Windows

  Quill

  Proofreads and rewrites your text in any Windows app, with a local model.

  A writing assistant that floats over any Windows app: select text anywhere, press a shortcut and get a word-level diff of corrections; Replace pastes back over the live selection. Runs on a local model or ChatGPT, and builds a personal writing profile across 19 error categories with drills and a stats dashboard.

  ShippedDesktop app

  See the screens

• Windows

  MD Viewer

  Read Markdown instantly and compare two documents side by side.

  A fast Markdown reader that also diffs two documents: Markdown, JSON, PDF and DOCX side by side, with tabs, an edit mode with live preview, session restore and Explorer right-click integration. Built on Electron; cold start around 390 ms.

  ShippedDesktop app

  See the screens

• Windows

  Clock Manager

  Keep every timezone you work with on screen, in one floating widget.

  A floating, always-on-top world-clock widget: a tray control panel to add clocks, pick a base timezone, switch layouts, compare timezones and, optionally, show FX parity per clock.

  ShippedDesktop app

  See the screens

• Web (PWA) · iOS · Android

  TagYa

  Design and print labels on Niimbot thermal printers, from the browser or the phone.

  A label editor for Niimbot thermal printers, in Portuguese: compose labels from text, QR codes, barcodes, icons and shapes in millimetres, preview the exact monochrome output and print over Bluetooth. The native app wraps the editor with a BLE bridge.

  LiveWeb + mobile app

  LiveGitHubSee the screens

More

• CortesMestreDesktop app · Windows

  Plan the cuts for a woodworking project with the least waste.

  A cut-plan optimiser for woodworking: guillotine bin-packing of sheet goods with per-sheet SVG layouts, yield and cost, cut sequence, 1D linear cutting for mouldings, CSV import, PDF export and label printing to a Niimbot D110.

  PrototypeSee the screens
• InvoicingWeb app · Web

  Run a freelance business: contracts, FX, invoices and cash flow in one place.

  Single-user freelance invoicing: clients, contracts with business-day pro-rata, FX tracking with locked rates, numbered PT/EN PDF invoices, scheduled sends, cash flow and reports, AI ingestion of invoices and contracts, and an MCP server so an assistant can prepare, preview, confirm and send an invoice.

  Personal