Senior Full-Stack Engineer and AI Engineer. 14 years building web and mobile products with TypeScript, React, React Native, Next.js, Node.js, Go and Python, five of them on Coinbase Wallet in fintech and crypto. Now I build LLM applications and AI agents with Claude, Claude Code and OpenAI: agentic workflows, multi-agent orchestration, RAG, evals and LLM observability.

Senior full-stack engineer with 14 years of experience shipping web and mobile products. I spent the last five years on Coinbase Wallet (web and mobile), building crypto purchase flows that move real money for non-technical people. That work put clarity, trust, and visual polish above everything else.

I work across the stack: React, React Native, TypeScript, and Next.js on the front, Node.js, Go, and Python on the back.

• Based in Paraná, Brazil. Portuguese (native), English (fluent).
• I care about technical writing and decision documentation, framing trade-offs so non-engineers can weigh in.
• Comfortable owning a feature end to end, from design-system components to backend services.

I build with the modern AI stack: Anthropic Claude and OpenAI, RAG, agent design, prompt engineering, evals, and LLM observability. Most of that work is public in Fhorja, an agentic workflow for AI coding agents such as Claude Code, Cursor and Codex. What it shows:

• Agent design and prompt engineering. 98 commands that an AI coding agent runs as Agent Skills, all sharing one output contract, taking a task from plan to a draft pull request.
• Multi-agent orchestration. Independent slices of work run in parallel sub-agents, each in its own git worktree, with a build, typecheck and test gate after every wave.
• Model routing. Sub-agents are routed by role across Claude models: mechanical work goes to a faster model, plans and reviews to a stronger one.
• Evals. 144 regression scenarios for the command outputs, and trigger evals on 68 of the 98 commands.
• Cost and quality measurement. Token use is recorded per task. In one experiment, batching about five small items per worker cut a 20-item fan-out from 17.69 USD to 9.58 USD.

Fhorja is an open-source agentic workflow for AI coding agents that I design and build solo, MIT licensed, at v2.1.0. An agent takes one task to a draft pull request: it plans, records its decisions on disk, runs independent work in parallel sub-agents, and lists what it decided or could not verify, so a person reviews and merges. In one measured run, three parallel sub-agents finished a wave in about 62 s instead of 173.6 s in sequence. Those slices took about a minute each, and each wave adds about 45 s of fixed cost. Source: github.com/Mozurok/fhorja.dev.

Let's talk about full-stack, mobile or AI engineering: LinkedIn, email (bruno.mazurok.c@gmail.com), or fhorja.dev.