João Carreiro
- Agents are dead. Long live the system.
- Context is dead. Long live Dreaming.
The model is dead. Long live the harness.
I define problems, design solutions, challenge the architecture and validate the outcome, while AI handles the implementation. Like a chef running a Michelin-star kitchen.
Contact meNo boring stuff. Impossible is welcome.
Is this the new CV?
I asked ChatGPT where I rank among its users, and to be honest about it.
Here’s the usual CV ↗, if you fancy those.
- Systems thinkingTop 1%
- Breadth and learning speedTop 1%
- Precision and methodical executionTop 5–10%
- Technical problem solvingTop 1–2%
- Spotting inconsistenciesTop 1–2%
- Product and strategic reasoningTop 2%
The good “…high general reasoning + unusually high agency + very broad transfer learning.”
The bad “…fairly low tolerance for tedious verification 😄.”
Read the full chat ↗I promise I write better English when talking to real people.
The Roads Not Taken
Current research focus
What if an AI-assisted development run could show the important decisions behind one solution, the context-dependent alternatives we did not take, and eventually let us explore a different path?
Decision Multiverse aims to map consequential choices without preset branch counts or categories. First we will test whether we can capture decisions faithfully, distinguish evidence from inference, and present a useful visual map. Later we hope to replay selected alternatives in isolation and compare real solution quality, cost and human intervention. Unexecuted paths are hypotheses, never observed results.
Core question: does exploring alternative decisions deliver better solutions than following one path with the same total resources?
Flagships
Where most of my time goes
Shopping AssistantPrepare your household’s weekly Continente cart from its shopping habits, compare alternatives and review everything before completing the order.Beta
Beta
Prepare your household’s weekly Continente cart from its shopping habits, compare alternatives and review everything before completing the order.
31.6k lines · 413 commits
CapacityShare how much capacity you have right now, with one person or a small group, without turning care into another conversation.Beta
Beta
Share how much capacity you have right now, with one person or a small group, without turning care into another conversation.
21.4k lines · 166 commits
Development CycleA control plane for supervised AI-assisted development, with pluggable runners, durable state, validation gates and correction loops.Creator
Creator
A control plane for supervised AI-assisted development, with pluggable runners, durable state, validation gates and correction loops.
15.6k lines · 177 commits
AI Server CommanderGive ChatGPT and Claude chat sessions terminal access to your server, without separate API, Codex, or Claude Code credits.Creator
Creator
Give ChatGPT and Claude chat sessions terminal access to your server, without separate API, Codex, or Claude Code credits.
7.2k lines added · 76 commits
Track StudioYour car's data on top of your track-day footage: speed, lap times, G-forces, pedals and the track map. Tesla Track Mode recordings are imported and synced automatically.Live
Live
Your car's data on top of your track-day footage: speed, lap times, G-forces, pedals and the track map. Tesla Track Mode recordings are imported and synced automatically.
11.3k lines · 203 commits
RadarCurated news and competitive intelligence for people who want signal, context and daily rhythm without drowning in noise.Beta
Beta
Curated news and competitive intelligence for people who want signal, context and daily rhythm without drowning in noise.
1,017 editions since Mar 2026
Kids ExplorerA deliberately simple voice-and-camera experience for the Rabbit R1 that lets a young child ask questions and explore objects around them. Install it with your own OpenAI key.Beta
Beta
A deliberately simple voice-and-camera experience for the Rabbit R1 that lets a young child ask questions and explore objects around them. Install it with your own OpenAI key.
4.0k lines
Shipped
Mine, live or in beta
Octopus Role BenchmarksA role-based LLM benchmark that turns public quality, coding and pricing data into practical model choices for multi-agent systems.Creator
Limited-Time DealsTemporary retail opportunities presented like a proper online shop. Works with Lidl and Aldi, with stock signals and alerts instead of messy leaflets.Live
Live
Temporary retail opportunities presented like a proper online shop. Works with Lidl and Aldi, with stock signals and alerts instead of messy leaflets.
8.4k lines · 65 commits
Refurb RadarFind refurbished products that are actually good deals by comparing them with validated new-market prices, condition and warranty.Beta
Beta
Find refurbished products that are actually good deals by comparing them with validated new-market prices, condition and warranty.
5.8k lines · 51 commits
OpenAI AdapterA generic adapter that lets OpenAI-compatible models use tools through a common command contract, without patching every host project.Creator
Creator
A generic adapter that lets OpenAI-compatible models use tools through a common command contract, without patching every host project.
5.1k lines · 31 commits
Headful Auth TunnelA tiny browser tunnel for AI agents: when automation hits a login, 2FA or CAPTCHA, a human solves it and the agent carries on.Creator
Creator
A tiny browser tunnel for AI agents: when automation hits a login, 2FA or CAPTCHA, a human solves it and the agent carries on.
4.2k lines · 23 commits
Super SearchA metasearch tool for web discovery, people lookups and vehicle/plate discovery, with clear engine selection and deep-search options.Live
Live
A metasearch tool for web discovery, people lookups and vehicle/plate discovery, with clear engine selection and deep-search options.
1.4k lines
Server ObservabilityA compact Linux server and network monitoring stack that catches storage stalls, blocked processes and network failures early.Creator
Creator
A compact Linux server and network monitoring stack that catches storage stalls, blocked processes and network failures early.
1.1k lines
Contributions
Open source I helped build
OctopusSecond-largest contributor, with 89 merged pull requests: hardened how multi-agent runs survive failure, with process supervision, retries, recovery after interrupted runs and quality gates.#2 contributor
#2 contributor
Second-largest contributor, with 89 merged pull requests: hardened how multi-agent runs survive failure, with process supervision, retries, recovery after interrupted runs and quality gates.
90 merged PRs
ClawSweeperContributed the repair flow that takes a failed PR to safe branch edits, with planning, validation and checkpoint gates.Contributor
Contributor
Contributed the repair flow that takes a failed PR to safe branch edits, with planning, validation and checkpoint gates.
5 merged PRs
OpenVNCViewerA VNC viewer for Windows that also logs in to macOS Screen Sharing. Contributed account credentials limited to ARD, isolation of events from replaced sessions, remembered window preferences and text-send progress with cancel.Contributor
Contributor
A VNC viewer for Windows that also logs in to macOS Screen Sharing. Contributed account credentials limited to ARD, isolation of events from replaced sessions, remembered window preferences and text-send progress with cancel.
7 merged PRs
In the works
Being built, or running only for me
Regime SuiteTradingView indicators in Pine Script that classify the market as trend or rotation from range, Value Area and momentum readings, with confirmation streaks and debug labels that show what each decision saw.Private
Private
TradingView indicators in Pine Script that classify the market as trend or rotation from range, Value Area and momentum readings, with confirmation streaks and debug labels that show what each decision saw.
5.0k lines · 75 commits
Freqtrade StrategiesThe Python side of Regime Suite: Freqtrade strategies that share its regime gate, FreqAI models and automated rolling hyperopt runs. Upstream, two merged fixes to Freqtrade’s profit reporting.Private
Private
The Python side of Regime Suite: Freqtrade strategies that share its regime gate, FreqAI models and automated rolling hyperopt runs. Upstream, two merged fixes to Freqtrade’s profit reporting.
2 merged PRs
Trend SuiteThe trend side of Regime Suite, grown from my earlier Hull Trading Suite: a TradingView indicator in Pine Script that enters and exits on Hull trend signals, filtered by trend strength, volume, Stochastic RSI and candle body.Private
Private
The trend side of Regime Suite, grown from my earlier Hull Trading Suite: a TradingView indicator in Pine Script that enters and exits on Hull trend signals, filtered by trend strength, volume, Stochastic RSI and candle body.
TradingView WorkbenchTradingView desktop bridge for inspecting live charts, Pine scripts and debugging indicators through controlled MCP tooling.Private
Private
TradingView desktop bridge for inspecting live charts, Pine scripts and debugging indicators through controlled MCP tooling.
5.0k lines · 75 commits
Memory KeeperA family memory app that preserves stories from parents and grandparents, in their own words, through guided voice and video.Development
Development
A family memory app that preserves stories from parents and grandparents, in their own words, through guided voice and video.
4.5k lines
CompraGuardA private purchase-protection app designed to turn receipts and invoices into useful records, deadlines, reminders and next actions.Validation
Validation
A private purchase-protection app designed to turn receipts and invoices into useful records, deadlines, reminders and next actions.
Social PublisherA planned editorial workflow for personal X and LinkedIn posts, from discovering ideas and checking sources to original commentary and human review.Planning
Planning
A planned editorial workflow for personal X and LinkedIn posts, from discovering ideas and checking sources to original commentary and human review.
560 lines
Chief of StaffA private executive-operations layer for planning, follow-ups and recurring life/admin workflows.Private
Private
A private executive-operations layer for planning, follow-ups and recurring life/admin workflows.