AI Engineer — multi-agent orchestration, MCP servers, and the production plumbing underneath. Open source, runnable, documented.

📍 Limerick, Ireland · Portfolio · LinkedIn · Upwork · Email

I design systems where several models, tools and services cooperate on one real problem — agent councils that propose, debate and adversarially verify; MCP servers that give an LLM safe hands on real APIs; and the unglamorous layer underneath (multi-tenant data, auth, billing, deploys) that a real product needs to exist.

Most product source is private — it's commercial. Everything below is open, runnable, or written up as a case study you can read end to end.

_{A live card rendered from my own coding hooks. Built with devcard — MIT, works with any agent that commits.}

When a pull request arrives from someone you don't know — a person, or increasingly an AI agent — trust rests on a username and reading the diff by eye. SCPE adds a signed envelope proving who produced a contribution and that nothing was tampered with, verified offline with no protocol server and no new accounts, using signing keys the contributor's git host already publishes.

I wrote the spec and three independent verifiers — Python, Go and Rust — held to the same verdict across 18 normative test vectors by a differential test, plus 8 adversarial ones. Ships as a GitHub Action that seals or gates pull requests.

The protocol runs on itself: this repo and seven others fail a pull request that arrives with no AI-use disclosure — including the ones my own agents open.

Apache-2.0 → Site · Repo · PyPI

Fuses network scan, WiFi CSI, camera pose and mmWave radar into one explainable room-occupancy state: confidence = strength (trust weight × the source's own confidence × freshness decay), and the dashboard always shows why it believes what it believes.

Ships its own MCP server, so an LLM agent can ask the house what's happening. Loopback-only by default, camera frames never stored, and every sensor path mock-tested — camera, mmWave, WiFi CSI, BLE, and the whole network-discovery stack. Runs with no hardware at all: off-localhost the frontend self-switches to a simulator.

Python 3.11 · FastAPI · AGPL-3.0 → Repo

An alternative to the hosted Deputy connectors where your token never leaves your machine. Three auth modes (API token, OAuth 2.0 with loopback flow, iCal fallback) and self-service /my/* endpoints, so a plain employee token works without manager permissions — with graceful degradation when access is denied.

Tested end to end against a mocked Deputy API; writes stay locked behind an opt-in flag.

Python · FastMCP → Repo

Four lenses on a question, five on a repo (Architect, QA, Security, Docs-DX, Maintainer), propose independently and debate; then the pipeline computes convergence — which claims multiple lenses reached alone — adversarially attacks the top claims with dedicated refuters, and only lets a neutral chair synthesize from what survived.

Runs against any repo or question, heterogeneous models per lens, and a deterministic mock backend runs the full pipeline offline. Source private for now — happy to walk through it.

Public-domain text in, print-ready book out. A headless Chromium is the typesetter: Paged.js over a stylesheet produces a 6×9 PDF with running heads and real page breaks, alongside EPUB3/Kindle and deterministic generative cover art — the same book built from one source in six languages.

The pipeline and a Part I sample are public; the full adapted prose is a separate paid edition.

→ Repo

Two full apps on one backend: Ownly (owner dashboard — shifts, payroll, stock, cash-up, live orders board) and Tillr (customer ordering PWA with real-time tracking). Postgres Row-Level Security as the isolation boundary, Stripe Connect wired for split payments, Cloudflare at the edge.

First end-to-end version built solo in under two months (June–July 2026); still under active development. Two demo tenants run on the same schema a real store would get, and the card rail is proven end to end in Stripe live mode — infrastructure proven, not a customer: no restaurant has run on it.

→ Case study · Ownly demo · Tillr demo

A white-label app for door-to-door reps. Its core is Wingman, a retrieval-augmented copilot (Supabase pgvector + Gemini) that answers a rep's question from the product's own playbook — grounded and cited.

The case study is the honest version: an IVFFlat recall collapse on a tiny corpus, a thinking-token budget eating the output, and a cost-ordered model fallback under quota. What actually broke, and how I fixed it.

→ Case study · Live demo

A minimal, clone-and-run RAG pipeline: chunk → embed → pgvector → grounded answer that refuses to invent. docker compose up, a local sentence-transformers model for keyless indexing, pytest, and a match_chunks() SQL function shaped like a Supabase RPC. The inspectable code behind my copilots.

→ Repo

The card at the top of this page. A local hook captures real AI-assisted coding output (not editor time) and syncs an anonymized copy — language and line counts, never project names or code — to a Cloudflare Worker + D1 that renders an embeddable SVG: i18n by viewer language, auto dark/light, pinned repos with live stars.

It hooks git itself, so it works with Claude Code, Codex, Cursor, local models — anything that commits. One-command setup.

MIT → Repo

_{Also: scrcpy-tray — your Android phone in the Windows tray, one click to mirror. MIT.}

Also working with: multi-agent orchestration · MCP (Model Context Protocol) · RAG pipelines (pgvector, sentence-transformers) · Postgres Row-Level Security · Supabase Edge Functions · Stripe Connect · agent hooks and CI-gated agent workflows

Open to freelance work and AI-engineering roles — multi-agent systems, MCP servers, RAG pipelines, and full-stack products built to ship.

• 🌐 Portfolio & CV — augustobastos.pages.dev
• 💼 LinkedIn — linkedin.com/in/augustobastos
• 🟢 Upwork — freelance profile
• ✉️ Email — augustobastos123@gmail.com