Bruno Brito

AI-powered satirical news platform — the project where I first wired generative AI into a full product: real headlines in, generated satire and cartoon illustrations out. Two independent services (Kotlin/Spring Boot backend, Next.js 14 frontend) share a MongoDB; cartoon images are produced at scale via the OpenAI Batch API; countries are tenant-scoped by subdomain. Its architecture became the blueprint for Recipe Repository.

Currently offline for maintenance — happy to show the code on request.

Full-stack recipe discovery platform. Users pick three to five ingredients and get matching recipes — and when nothing matches, a multi-agent AI pipeline generates a new recipe, images, and enriched content on demand. Content is localized per locale via URL subdomains, with SSR for SEO.

The AI layer is the interesting part: multiple AI tools interact in a pipeline — one model generates content, others enrich it, generate images, and a sanitizer validates user-entered content before it is ever stored. I designed the architecture so each AI concern is an isolated, testable service.

Information website about solar eclipses around the world — per-city data on what's visible, when, and how, plus maps and sky previews. Built in Next.js with on-demand per-city SSR (each city is a deep-linkable, SEO-indexable page) and a lean Docker deployment I scripted end-to-end, including a fast deploy path that hot-swaps build artifacts into a running container with BUILD_ID verification and health checks.

Swing-trading assistant that turns raw market data into actionable, explainable signals. A scheduled pipeline scans EU and US markets: OHLCV from Tiingo into a DuckDB cache, rule-based signals (RSI, MACD, MA crossovers, volume spikes) blended with a GradientBoosting ML model and an ensemble meta-model into a single conviction score. A breaking-news monitor scores headlines with a neural embedder and triggers intraday re-scans.

Private project. The part I'm proudest of: a layered signal architecture (rules → ML → ensemble) where every recommendation decomposes into the factors that produced it — and a mobile action-button loop (Enter / Skip / Confirm exit) that closes the position-tracking cycle.

A personal AI-assisted development pipeline I've hardened into a repeatable practice: agentic coding agents with specialized roles (planning, design, senior review) working against a shared task spec, where the human stays the architect and the agents do the heavy lifting. High-volume, low-stakes tasks are routed to local LLMs to keep cloud spend flat, while judgment calls and production work stay on strong models — with read-only tooling in production so an agent can never break what's live.

The same discipline I run at work: AI as a measured productivity layer — with cost, safety, and code quality in the loop — never a black box. I can walk through the setup, the trade-offs, and the cost numbers on request.