Yago Lagrotti BraccoSenior Software Engineer

The Challenge

Ingest ~3 million daily events from Meta, Google, and TikTok APIs into a multi-tenant platform while keeping OLTP latency under control and real-time dashboards responsive without bottlenecks.

The Architecture

I led the redesign from a Node.js monolith to an event-driven microservices architecture. Kafka decouples ingestion from processing; Golang consumers handle stream processing at scale; ClickHouse absorbs analytics queries so PostgreSQL handles only transactional workloads. Strangler Fig pattern for zero-downtime migration.

The Challenge

Selling a part nobody has seen yet. The customer must trust the result before production starts — and once printed, there is no undo. Add an order that crosses quoting, artwork, payment, production and delivery, each with its own rules, and the real risk stops being technical and becomes commercial: printing the wrong thing costs material and time.

The Architecture

A 3D visual proof generated automatically from the uploaded file and rendered in the browser, which the customer approves before anything reaches the printer — approval is an explicit gate in the flow, not a notice. Underneath, a modular monolith of twenty bounded contexts where module boundaries are enforced by tooling, not by convention.

The Challenge

A security tool that gets in the way gets abandoned: if running the project becomes harder than it was with a .env, the team goes back to the .env. Each protection had to cost almost nothing day to day and still keep the value out of the repository, the terminal, the logs and an AI agent's context. The next step is harder still: sharing secrets with a team through a server without having to trust it.

The Architecture

I built a Go CLI on my own, shipped as a single binary. Values live in a local vault encrypted with XChaCha20-Poly1305, its key protected by the master password through Argon2id; the project keeps only names, in a versionable envrune.yml; and envrune run injects the variables straight into the process and masks its output. For teams, EnvRune Cloud is optional and end-to-end encrypted: the server stores ciphertext and names and, by default, cannot read a single value.

The Challenge

Presenting the same tree as a 2D map, list and 3D constellation requires consistent hierarchy and progress. Collaborative editing must respect permissions and prevent invalid node relationships.

The Architecture

React Flow renders the map and on-demand Three.js presents the constellation. Supabase provides authentication, PostgreSQL with RLS and Realtime synchronization. The product includes an optional AI coach, Stripe subscriptions and PWA installation.

The Challenge

Stores silently lose revenue: stockouts, payment errors and API outages go undetected for hours. The problem is not the store breaking — it is nobody finding out while revenue drains.

The Architecture

I built the whole platform from scratch as sole engineer: configurable health-check workers, a threshold-based alerting engine and a multi-channel notification pipeline (email, SMS, Slack), with a Node.js/TypeScript API over PostgreSQL and a real-time dashboard.

The Challenge

Print products have no fixed price: it varies by material, finish, dimensions and quantity, and some combinations only make sense as a custom quote. At the same time, printing is irreversible — wrong artwork approved by mistake becomes material loss, not a bug you patch in production.

The Architecture

A pricing engine over configurable options that calculates on the server and recalculates before writing the cart, with an escape hatch into custom quoting when a product has no fixed price. And an artwork flow with explicit approval, which only then releases production.

The Challenge

The process is complex and error-prone, and has historically been run on spreadsheets and email chains. The hard part is not storing the document: it is guaranteeing the dossier is complete and consistent for every SKU — and being able to prove it.

The Architecture

I designed and shipped the entire product as sole engineer: a NestJS API with an automated dossier generation pipeline, AI-assisted document creation, structured data validation and a Next.js dashboard for compliance teams.

The Challenge

An architecture project becomes a pile of scattered files: plans, technical drawings, references, site photos and decisions agreed over WhatsApp. The client does not know where to look, and the studio repeats the same information for months. The problem was not storing files — it was giving them a place where they make sense together.

The Architecture

The floor plan became the interface. The architect drops a pin on each room and accumulates as many drawings, images and notes inside it as needed — the plan stops being a picture and becomes the project's index. Around it, a portal where the client talks to the studio and downloads documents, and a panel where the studio publishes what it wants shown on the site.

The Challenge

Most of the usual conversion levers are off the table. Brazil's psychology council forbids promising outcomes and publishing patient testimonials — precisely what a service site normally leans on. On top of that, what flows through it is sensitive health data under Brazil's data-protection law. The site had to convert through clarity, not persuasion.

The Architecture

Trust was built from what is allowed and verifiable: the professional licence number always visible, an honest explanation of how therapy works, frequently asked questions answered plainly, and a short booking flow. Reach came from solid technical SEO, not from advertising.

The Challenge

Removing the human bottleneck on that kind of task without giving up review. An agent writing straight to the main branch is a liability; the value only shows up if it enters through the same quality gate as anyone else on the team.

The Architecture

The agent acts as a virtual developer inside the workflow that already exists: it pulls the card, interprets the requirement, writes the implementation, runs the tests and opens a pull request. Human review remains the final step — what changes is who writes the first draft.