AI Engineer · Design Engineer

I design the part of AI people have to trust.

Chat is the easy part. I build the approval card, the refusal, the citation, the cost meter, and the validation layer underneath that keeps the model honest.

See the work Résumé

15+: years shipping
5: MCP servers built
3,700+: users on my SaaS
1M+: behind a login I led

agent.run · outreachdemo · scripted run

cost $0.00000 tok0.0s

What I do

Two halves of the same job.

Agents and MCP

Agents with tool calling, MCP servers with real auth, retrieval that knows each user. Explicit orchestration, so every tool call stays inspectable.

Interfaces people trust

The screens after the chat: approvals for irreversible actions, refusal states, citations, cost, diffs to review. Fifteen years of frontend in production behind them.

The layer underneath

Model output validated like user input: schema, field assertions, guardrails that can refuse, and failure that is explicit instead of silent.

Selected work

Products I designed, built and run.

All work

ViralRecast

Live since Sep 2026

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

SaaS
MCP
pgvector
Stripe

Renderster

In production

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

Go
Redis queue
FFmpeg
API

Mentezinha

In production

An AI for kids that knows when to say no.

Vertex AI
RAG
Guardrails

StartAtende

In production

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

WhatsApp
Agents
MCP

SunoBoard

In production

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

AI music
MCP
npm

Linkezera

3,700+ users

One link for everything, with analytics, quizzes and leads. 3,700+ people use it.

SaaS
Analytics
MCP

Model Context Protocol

Five MCP servers. Your assistant can already use them.

I wrote the book on MCP in Portuguese, then shipped these to production, which is a very different experience.

44 tools

read_postViralRecast search_viral_contentViralRecast trending_nowViralRecast similar_contentViralRecast analyze_creatorViralRecast search_adsViralRecast advertiser_adsViralRecast read_adViralRecast transcribe_videoViralRecast download_videoViralRecast create_clipsViralRecast get_jobViralRecast my_historyViralRecast get_accountViralRecast generate_musicsuno-mcp wait_for_musicsuno-mcp get_music_statussuno-mcp list_recent_musicsuno-mcp get_creditssuno-mcp list_quizzesLinkezera get_quizLinkezera get_quiz_responsesLinkezera get_leadsLinkezera read_postViralRecast search_viral_contentViralRecast trending_nowViralRecast similar_contentViralRecast analyze_creatorViralRecast search_adsViralRecast advertiser_adsViralRecast read_adViralRecast transcribe_videoViralRecast download_videoViralRecast create_clipsViralRecast get_jobViralRecast my_historyViralRecast get_accountViralRecast generate_musicsuno-mcp wait_for_musicsuno-mcp get_music_statussuno-mcp list_recent_musicsuno-mcp get_creditssuno-mcp list_quizzesLinkezera get_quizLinkezera get_quiz_responsesLinkezera get_leadsLinkezera

get_overviewStartAtende get_agent_performanceStartAtende get_business_insightsStartAtende get_faq_gapsStartAtende get_model_costsStartAtende get_sentiment_trendStartAtende get_top_intentsStartAtende get_top_clientsStartAtende get_productsStartAtende get_payment_summaryStartAtende list_conversationsStartAtende ollos_probeollos-mcp ollos_transcribeollos-mcp ollos_diarizeollos-mcp ollos_keyframesollos-mcp ollos_framesollos-mcp ollos_read_screenollos-mcp ollos_searchollos-mcp ollos_reviewollos-mcp ollos_jobollos-mcp ollos_cancelollos-mcp get_overviewStartAtende get_agent_performanceStartAtende get_business_insightsStartAtende get_faq_gapsStartAtende get_model_costsStartAtende get_sentiment_trendStartAtende get_top_intentsStartAtende get_top_clientsStartAtende get_productsStartAtende get_payment_summaryStartAtende list_conversationsStartAtende ollos_probeollos-mcp ollos_transcribeollos-mcp ollos_diarizeollos-mcp ollos_keyframesollos-mcp ollos_framesollos-mcp ollos_read_screenollos-mcp ollos_searchollos-mcp ollos_reviewollos-mcp ollos_jobollos-mcp ollos_cancelollos-mcp

How I build

An agent does not fail loudly.

It returns something plausible, the logs look clean, and weeks later someone finds a corrupted record. So these are the rules.

01
Validate output like user input
A successful API call says nothing about whether the payload makes sense.
02
Refusal is a feature
When the answer is not in the source, "no" is the correct output. Design that screen.
03
Deterministic first
Checks that cost nothing run before any model judges anything.
04
Cost per conversation
Not per call. A conversation is what a business can price against.
05
Humans approve the irreversible
No agent writes to a CRM, sends outreach or spends budget without passing a gate.
06
Traceability is not verification
Citing a source is not the same as matching it. Numbers get checked against the source.

Off hours

I also make games.

Five published on itch.io, scored with an AI music studio I built for that reason.

Pick a game

Ask the machine

An agent that cannot make things up.

ASK-1 answers about my work by quoting its sources, never by generating text. Every answer shows where it came from, and when the answer is not there, it says so. No model call, no API key on the server, nothing to abuse.

ASK-1 retrieve → rank → quote → cite

Ask anything about my projects, stack or experience. Pick a suggestion to start.