UX, UI and AI · React · Next.js · TypeScript · Design Systems · Accessibility · MCP · AI Agents · LLM Evaluation 15+ years shipping production software · Remote, Brazil (UTC-3) · kelvinbiffi.com

AI made me faster. It did not make me necessary. Fifteen years of engineering did that, and it is why I can tell a model where to look instead of watching it guess.

I wrote the books on MCP, Context Engineering and AI Agents in Python. Then I shipped all three to production, which is a very different experience.

An agent does not fail loudly.

It returns something plausible. The logs look clean. Three weeks later someone finds a corrupted record.

So I stopped trusting model output and started validating it like user input: schema-constrained, value-checked, with guardrails that can refuse to answer, evaluation loops, and cost tracked per conversation instead of per call.

That layer is half of what I do. The other half is the interface on top of it, and it is the half almost everyone underbuilds.

Chat is the easy part. The hard part is every screen after it: the approval card for an action that cannot be undone, the refusal state where the model said no and that is the correct answer, the tool call you can inspect, the citation that ties a claim back to its source, the cost meter, the diff someone has to review before a record changes. Those screens decide whether a person trusts the system enough to use it. A model’s answer is worthless until someone can act on it.

Games are why I built SunoBoard in the first place. Under K3 Games Studio I design, build and publish my own, and five are live on itch.io:

• Subscription, billing and lifecycle flows on Stripe, used by thousands of paying members at Firstleaf (US wine subscription)
• An enterprise login and registration gateway with Keycloak and MFA serving millions of users at Lojas Renner
• Frontend systems for DIFC, the Dubai International Financial Centre, and for 605 Media Analytics
• A Vue.js landing page builder at Pmweb that cut development effort by 70%, moving page creation out of the engineering queue and into the client's hands
• Near Companion at Hire With Near: a Chrome extension over an LLM pipeline with Recall.ai, Supabase, HubSpot, Cloudflare R2 and n8n. A bot joins the intake call as a silent participant and writes the structured job description live, during the call. I owned the whole interface, and every model output was schema-validated before it touched a CRM

Surface

AI in production

Underneath

Agents with tool calling · MCP servers and clients · RAG and contextual retrieval · evaluation loops · guardrails · schema-constrained output · cost control · FFmpeg · WhisperX

How I work: Claude Code daily, with a written context file in every repo so the agent inherits the project's decisions instead of rediscovering them every session.

Client work lives in private repos. What is open is everything I build under YellowKode and Wunka, both mine, plus the tooling here.

The official clients for Couryo, open source under Wunka. Each one wraps the API, retries safely with an idempotency key, and plugs into the mail layer the framework already has, so switching providers is a config change and not a rewrite.

• couryo-node - TypeScript SDK. On npm as @couryo/node.
• couryo-python - Python SDK with a Django email backend. On PyPI as couryo.
• couryo-php - PHP SDK with a Laravel mail transport. On Packagist as couryo/couryo-php.
• couryo-email-templates - Brazilian transactional email templates for React Email (Pix, boleto, password reset). Works with any provider. On npm as @couryo/email-templates.

• ollos-mcp - eyes and ears for agents. Whisper transcription, speaker segmentation, perceptual-hash keyframing and OCR, all running locally in Node through ONNX Runtime. No Python, no cloud, no API key, so the file never leaves the machine. Same core ships as an MCP server, a CLI and a Node library. The part I care about is that it reports confidence per block and per segment and filters the lines Whisper invents over silence, because a transcript that does not say where it is unsure is worse than one that is visibly wrong. On npm as ollos-mcp.
• suno-mcp - MCP server that lets Claude, Cursor and any MCP client generate music with Suno from a plain sentence. It owns the whole generation flow, polling and retrieval, so the assistant never sits blocked on the API. On npm as @yellowkode/suno-mcp, with suno-mcp-core holding the typed handlers and tool schemas so a new transport never means rewriting the tools.
• yk-agent - autonomous agent that runs entirely on your machine. Browses, extracts and executes with no API key and no account, with a retrieval knowledge base built in. The interesting constraint is privacy: nothing leaves the laptop, which changes what you are allowed to do at every step.
• yk-chat - local AI chat with the same knowledge base, for people who want to use a model without handing over their data.
• squadron - an opinionated skeleton for a project run by one person and a team of agents. Ten agents, five skills, and documentation the agents actually obey. Not a framework, a starting point you adapt.

• Guide-Book - design system and style guide generator that reads your CSS, SCSS, LESS or Stylus and produces the categories, elements and modules of your design system. On npm as guide-book-generator.
• javascript-book - every code example from JavaScript: Básico ao Avançado.
• Book source, all public: MCP na Prática com Python, Agentes de IA com Python and Multiagente com Python. I wrote the books, then shipped the same patterns to production, which is a very different experience.
• Bubble-Shooter and GUI_Unity3D_Style_Guide_Base - a JS game and a Unity3D GUI style guide. Games are the reason I built SunoBoard (the published ones are above), and a design system is a design system whether it renders in a browser or in Unity.
• Criando e aplicando um Style Guide do zero - full source for my design systems course.

Almost thirty published books. Four of them are the exact stack I work in.

AI engineering

Remote, Brazil (UTC-3), so most of a US or European day is my day too.

kelvinbiffi.developer@gmail.com · LinkedIn · kelvinbiffi.com