Every machine, as a list
Skip to the résumé

PLAYER 1 · AI ENGINEER

KELVIN BIFFI

I build the part of an AI product that people see and need to trust.

Chat is the easy part. I build the screens after it, and the layer underneath that keeps the model honest.

The arcade

Walk with ← → or A D. Press Enter at a machine. Or just click one.

Every machine, as a list

High scores

  1. 1ST 15+ years shipping production software
  2. 2ND 3,700+ users on Linkezera, a SaaS I built and run
  3. 3RD 1M+ users behind a Keycloak + MFA login I led
  4. 4TH 70% less dev effort with a landing page builder I built
  5. 5TH ~30 published books, four on MCP, agents and context
  6. 6TH 5 games published on itch.io

Player stats

Interface LV 10

React · Next.js · TypeScript · Tailwind · Design systems · Chrome extensions

AI in production LV 9

Agents · tool calling · MCP servers and clients · RAG · pgvector · Vertex AI · OpenAI · Gemini · guardrails · evals

Underneath LV 9

Node.js · Python · FastAPI · Go · PHP · PostgreSQL · GraphQL · Redis · Stripe

Automation LV 9

n8n · webhooks · FFmpeg · WhisperX · Recall.ai · HubSpot

Ship it LV 8

Docker · CI/CD · GitHub Actions · Playwright · Vitest · Cypress

Ask the machine

A small agent that answers questions about my work. It quotes its sources instead of generating text, so it cannot invent anything: every answer shows where it came from, and when the answer is not there, it refuses. No model call and no API key on the server.

ASK-1 retrieve → rank → quote → cite · no model
Model calls: 0 Cost: $0.00 0 questions

Shipped for other people

Client work lives in private repos. This is what it did.

I wrote the books, then shipped them

Almost thirty published books. These four are the exact stack I work in (in Portuguese).

CONTINUE?

Taking on remote contracts with US and European teams. UTC-3, so most of your day is my day too.

Email me Download résumé (PDF)

LinkedIn GitHub itch.io YellowKode kelvinbiffi.developer@gmail.com

SaaS · MCP server · pgvector

ViralRecast

Shows what is going viral, why it worked, and cuts the clips for you.

Live since Sep 2026

The problem

Creators and marketers lose hours scrolling to find what is working, and then still have to guess why it worked.

What I built

Finds what is going viral on TikTok, Reels and YouTube, explains the anatomy of each post (hook, emotion, replay peaks) and cuts the vertical clips with captions. A web dashboard for people, and an MCP server so the same product runs inside Claude and ChatGPT.

The decision that mattered

One product, two doors, one wallet. People use the dashboard, agents use MCP, and both spend from the same atomic credit wallet that refunds a job when it fails. The balance a user sees is always the balance they have.

Receipts

  • MCP server with OAuth 2.1, PKCE and dynamic client registration
  • Postgres 16 with pgvector and an HNSW index for post similarity
  • Atomic credit wallet with automatic refund of failed jobs, Stripe billing

Stack React 18 · TypeScript · Vite · Tailwind · shadcn · PHP 8.3 · Postgres 16 · pgvector · Redis · OpenRouter · Stripe

Visit viralrecast.com →

Video rendering API · Go

Renderster

Send JSON or use the editor. Get a finished video back. Built in Go.

In production

The problem

Rendering video from code usually means babysitting FFmpeg on a server that falls over under load and loses jobs when it does.

What I built

Video rendering as a service, and the media engine behind ViralRecast. A visual editor or a JSON spec goes in, a finished video comes out. Go API and workers on a Redis queue, FFmpeg underneath, output on Cloudflare R2.

The decision that mattered

The same render has two doors: a visual editor for people and JSON for machines. The API docs are generated from the code, including an llms.txt, so an agent can read the API without a human translating it.

Receipts

  • Go 1.23 backend with chi, pgx and go-redis
  • asynq queue with retry, backoff, visibility timeout and dead-letter
  • Concurrent FFmpeg workers with goroutines, output on Cloudflare R2

Stack Go 1.23 · chi · pgx · Redis · asynq · FFmpeg · Cloudflare R2 · React · Konva · Zustand · OpenAPI

Visit renderster.com →

Child-safe conversational AI

Mentezinha

An AI for kids that knows when to say no.

Vertex AI · RAG per child

The problem

With children as the users there is no acceptable failure rate, and "the model usually behaves" is not a safety plan.

What I built

I built it for my son's homeschooling, and now other families use it. Each child has a retrieval index on Google's managed vector search that grows beyond what the parent configured, accumulating understanding across conversations so the next one is better.

The decision that mattered

The guardrail is a validation layer with the authority to refuse, not a line in a prompt. A refusal is treated as a correct outcome, not an error, because with children it often is the right answer.

Receipts

  • RAG per child on Google Vertex AI with managed vector search
  • Personalization that accumulates across conversations, not generic document RAG
  • Guardrail that can refuse, enforced outside the prompt

Stack Google Vertex AI · managed vector search · RAG · guardrails

Visit mentezinha.app →

AI support on WhatsApp

StartAtende

A WhatsApp support agent that also tells the business what customers want.

Agent + product intelligence

The problem

A support agent that only answers throws away the most honest data a business has: what its customers actually say.

What I built

AI customer support on WhatsApp with intent routing. Then the agent's own conversations with users, customers and leads are analyzed for sentiment, product gaps, business opportunities and market signals. The agent does not just answer, it becomes a source of product decisions.

The decision that mattered

Cost is tracked per conversation, not per API call, because a conversation is the unit a business can actually price against.

Receipts

  • Intent routing on WhatsApp
  • Cost per conversation
  • Sentiment and product-gap analysis over the agent's own conversations

Stack LLM agents · WhatsApp · intent routing · sentiment analysis · MCP

Visit startatende.com →

Link-in-bio SaaS

Linkezera

One link for everything. 3,700+ people use it.

3,700+ users

The problem

Creators and small businesses in Latin America need one link that holds everything, in their language, at their price.

What I built

A link-in-bio SaaS built end to end and still run by me, with 3,700+ users across Latin America. It also exposes its quizzes and leads over MCP.

The decision that mattered

Running it myself is the point. 3,700 real users is where you learn which screen people actually use, and which one only looked good in the design file.

Receipts

  • 3,700+ users
  • Built and operated end to end
  • Quizzes and leads available over MCP

Stack SaaS · MCP

Visit linkezera.com →

AI music studio · MCP

SunoBoard

An AI music studio that an agent can run on its own.

Built to score my games

The problem

I wanted original soundtracks for the games I make, and generating music by hand in a web UI does not scale past a few tracks.

What I built

An AI music studio, plus an MCP server so an agent can compose, check generation status and pull finished tracks without a human in the loop. The server is open source on npm as @yellowkode/suno-mcp.

The decision that mattered

The MCP server owns the whole generation flow, polling and retrieval, so the assistant never sits blocked waiting on the API.

Receipts

  • Open-source MCP server on npm (@yellowkode/suno-mcp)
  • Typed handlers and tool schemas in a separate core package
  • The soundtracks of my games come from here

Stack TypeScript · Node.js · MCP · npm

Visit sunoboard.com →

Puzzle, arcade

Bubble Blaster

Match three, chain reactions, 24 neon levels.

Bubble Blaster

Aim, shoot, match three. Chain reactions drop whole clusters in slow motion across 24 handcrafted levels in 3 neon worlds, each with its own pixel art creatures and an original synthwave soundtrack.

Games are why I built SunoBoard in the first place. Under K3 Games Studio I design, build and publish my own.

Play →

Strategy

Kaiju Siege

Five mechs against the kaijus. Hold the city.

Kaiju Siege

The kaijus rose from the deep. Deploy five combat mechs, call orbital strikes and hold the city across four living biomes with day and night cycle and weather. Campaign with four bosses, plus an endless mode.

Games are why I built SunoBoard in the first place. Under K3 Games Studio I design, build and publish my own.

Play →

Action, arcade

Dino Breakout

The dinosaurs own the streets. Get the civilians out.

Dino Breakout

The fences failed at midnight and the dinosaurs own the streets. Lead the Rescue Rangers through the city and get the civilians out.

Games are why I built SunoBoard in the first place. Under K3 Games Studio I design, build and publish my own.

Play →

Action, horror

Hollow Fields

Harvest souls as Death's last servant.

Hollow Fields

The world did not end in fire, it rotted. You are Death's last servant, harvesting thousands of souls with auto-weapons, a manual reap sweep and a dash.

Games are why I built SunoBoard in the first place. Under K3 Games Studio I design, build and publish my own.

Play →

Shooter, arcade

Gate of the Dead

Arcade shooter.

Gate of the Dead

Arcade shooter.

Games are why I built SunoBoard in the first place. Under K3 Games Studio I design, build and publish my own.

Play →