Facet · Agents

Agents & Agent Systems

Agents that serve real people — and close the order.

A real agent isn't the one that impresses in a demo — it's the one that talks to a real customer, understands the garbled order, and survives Friday night. That's what I build: a multi-channel sales agent with human handoff, an attendant that closes the order and dispatches it to the kitchen, conversational AI with memory serving 1,000+ users. Anyone can build a pretty demo; a customer served at 11pm is a different category.

Multi-Channel (WhatsApp · IG · FB) Human Takeover Anti-Promise Detector Tool Use / Function Calling Per-User Memory Structured Outputs RAG · pgvector Voice (Whisper) · Vision Agents as Entities Living UI (SSE) Multi-Channel (WhatsApp · IG · FB) Human Takeover Anti-Promise Detector Tool Use / Function Calling Per-User Memory Structured Outputs RAG · pgvector Voice (Whisper) · Vision Agents as Entities Living UI (SSE)

Talk to me 📄 CV — AI Agent Systems (PDF)

0 Agent systems in production

0 Vik Nutri users in healthtech

0 Channels served: WhatsApp · Instagram · Facebook · app

0 Humans on call in Salgadelli's customer service

My arc in agents runs from enterprise scale to the shop counter. Giants puts sales agents on WhatsApp, Instagram, and Facebook with multi-LLM routing, a human-takeover protocol that hands the human the entire conversation — and an anti-promise detector that holds risky output before it reaches the lead. Vik Nutri took conversational AI with tool use, structured outputs, and per-user memory to 1,000+ healthtech users, from agent design to mobile UX.

The most honest test of an agent is running a physical business. Salgadelli's attendant answers the shop's business WhatsApp with the real menu and prices, builds the order in natural conversation, dispatches it to the kitchen — Kanban on the TV, with voice confirmation — and the customer is notified at every step. No recorded demo, no rehearsed happy path: there's a hungry customer on the other end.

And then there's the frontier I care about now: the agent as the fabric of the product, not a chat bolted to the side. In Boleiragem, the game's entities are conversable agents and the AI manager runs the entire club via tool calls — the thesis is 'everything is conversation'. In Massinha, the interface itself is agent output: a UI organism that remolds itself over SSE as the conversation evolves.

Architecture

┌────────────────────────────────────┐
│  WhatsApp · Instagram · FB · App   │
└──────────────────┬─────────────────┘
                   │
┌──────────────────▼─────────────────┐
│          AGENTE DE FRENTE          │
│  tools · memória · voz da marca    │
└──┬──────────┬───────────────┬──────┘
   │          │               │
┌──▼────┐  ┌──▼────┐   ┌──────▼──────┐
│ANTI-  │  │HUMAN  │   │   PEDIDO    │
│PROMISE│  │TAKE-  │   │ estruturado │
│GUARD  │  │OVER   │   │ (checkout)  │
└───┬───┘  └───┬───┘   └──────┬──────┘
    │          │              │  estado
    │          │              │  por etapa
┌───▼──────────▼──────────────▼──────┐
│ Cozinha (KDS) · CRM · Métricas     │
│ Notificação do cliente por etapa   │
└────────────────────────────────────┘

The anatomy of my service and sales agents: channels up front; one agent with tools, memory, and the brand's voice in the middle; and the outputs that matter — a human when it's needed, an order in the kitchen when it closes, a customer notified at every step.

Giants is the blueprint for the sales agent a company can unleash without fear: genuinely multi-channel (WhatsApp, Instagram, Facebook), routed across Claude, GPT, and Gemini per task, and fitted with the two valves that separate a product from a toy — structured handoff to a human with full context, and an anti-promise detector that stops the agent from promising what the business won't deliver.

Impact

Multi-LLM agents with per-channel, per-task routing
Human takeover with the whole conversation, not a lost summary
Anti-promise detector: risky output dies before it reaches the lead

Stack

Rails Multi-LLM (Claude · GPT · Gemini) WhatsApp · Instagram · Facebook Hotwire PostgreSQL

The agent with the most honest trial by fire in the portfolio: a hungry customer. It answers Salgadelli's business WhatsApp with the real menu and prices, builds the order in natural conversation, and closes the operational loop — the kitchen gets it on the TV's Kanban, confirms by voice, and the KitchenNotifier updates the customer at every step. Customer service as a system, not as a chat.

Impact

In production on the shop's business WhatsApp, with real customers
Conversation → order → kitchen (KDS) → notification, no human in the middle
Live menu and prices — the agent never invents an offer

Stack

Ruby on Rails 8 WhatsApp (Evolution API) Claude Hotwire (KDS on a TV)

A conversational agent at enterprise scale: 1,000+ healthtech users talking to an AI that remembers them. Tool use, structured outputs, and per-user memory across sessions — plus the responsibility of being a health product, where a careless answer is not an option. I took it from agent design to mobile UX, on top of the multi-LLM platform I designed from zero.

Impact

1,000+ users in enterprise production
Per-user memory across sessions — a conversation that continues, not restarts
Full stack: agent design, platform, and mobile UX

Stack

Ruby on Rails React OpenAI · Claude · Gemini OpenRouter · LiteLLM

Project discovery is a conversation — so it became an agent. AircTech runs software project scoping by talking: it extracts requirements with function calling and structured outputs, scores project viability, and manages the client lifecycle. The briefing form is dead; the agent asks what matters and returns structured scope.

Impact

Conversation becomes structured requirements via function calling
Automatic per-project viability scoring
Client lifecycle managed end-to-end by the agent

Stack

Rails 8 React Native (Expo) OpenAI Claude PostgreSQL

My personal agent workspace: multimodal input (audio, image, PDF), agents pre-configured per context, and recurring workflows triggered by cron, webhook, or email. And the feature that says the most about agents at home: a Kids mode with tuned guardrails — because an agent that lives with children needs real limits.

Impact

Recurring agents: cron, webhook, and email as triggers
Multimodal input in the same flow (audio · image · PDF)
Kids mode: guardrails as a family feature, not a checkbox

Stack

Rails 8 Hotwire · Tailwind v4 Solid Queue Whisper · Gemini Vision Railter (LLM gateway)

A multimodal personal assistant where extensibility is the product: it speaks via voice (Whisper), sees (Vision), and acts via tool calling — with the capability set defined in the database. An agent that gains a new skill with an INSERT, not a deploy: the design that later became the pattern in my bigger systems.

Impact

Voice + vision + text in the same agent
DB-driven capabilities: new skill without a deploy

Stack

Ruby on Rails Whisper GPT / Gemini Vision Tool Calling

The agent that only says what the collection supports. Script generation over proprietary catalogs with RAG and pgvector: vector retrieval finds the right data, the LLM composes — and the rule is fidelity to the catalog, not loose creativity. RAG here is a truth contract with the client, not a buzzword.

Impact

Answers anchored in a proprietary catalog via pgvector
Fidelity as a requirement: the agent never invents the collection

Stack

Ruby on Rails pgvector Embeddings PostgreSQL

The most radical thesis: in a game, EVERYTHING can be conversation. In Boleiragem, the AI manager runs the club via tool calls — lineups, training, negotiations — and the world's entities (staff, scouts, press) are conversable agents with their own permissions. The player doesn't press buttons: they talk to a society of agents inhabiting a shared living world.

Impact

AI manager operates the entire club via tool calls
A society of agents: game entities are conversable, with permission and mediation
One living MMO world — the agents keep acting without you

Stack

Ruby on Rails 8 Hotwire (SPA · zero reload) Solid Queue Claude (AI manager via tool calls) PostgreSQL

Also in: Apps

Live →

If the agent decides, why is the screen fixed? Massinha flips the hierarchy: the interface is agent output — a UI organism that remolds itself over real SSE streaming as the conversation and its decisions evolve. Built engine first, with an eval loop measuring behavior before any aesthetics. It's the agent moving out of the chat box.

Impact

UI that remolds itself in real time over SSE as the agent decides
Engine + eval loop before form — behavior measured, not guessed

Stack

Ruby on Rails 8 SSE (real streaming) Hotwire / Stimulus Claude

May 2023 – Jul 2025

Principal Software Engineer · AI Platform Lead

VIK (Remote)

Shipped Vik Nutri to 1,000+ users: a conversational agent with tool use, structured outputs, and per-user memory — full stack, from the agent to mobile UX
Designed the company's agent platform from zero: function calling, RAG over proprietary data, multi-LLM routing via OpenRouter / LiteLLM
Conversational AI in healthtech: the cost of a wrong answer is real — I designed the agent for that

Ruby on Rails React OpenAI · Claude · Gemini OpenRouter · LiteLLM Kafka · RabbitMQ BigQuery · Looker

🤖

AI Agent Systems

Multi-agent orchestration in production. Tool-call loops, sub-agents, autonomous Claude Code workers, MCP servers, context engineering.

Multi-Agent Coordination MCP (Model Context Protocol) Function Calling / Tool Use Claude Code SDK Sub-Agents & Hand-offs Context Window Management

🏗

Multi-LLM Orchestration

Routing, fallback, and cost-aware orchestration across providers. Streaming, structured outputs, retries, circuit breakers.

Claude (Sonnet · Opus) GPT-4o / GPT-5 Gemini Flash / Pro Llama · Mistral OpenRouter / LiteLLM SSE Streaming

🎬

Multimodal Pipelines

Video, audio, image with AI — from capture to insight, in batch or real-time. WebSocket streaming, 8+ tools in flight.

Video Processing (FFmpeg) Computer Vision (GPT / Gemini Vision) Speech-to-Text (Whisper) Image Generation (Gemini Flash) Real-Time Streaming WebSocket / SSE

Want an agent serving your customers tomorrow?

Sales, service, and product agents — multi-channel, with human handoff and real metrics. For teams hiring or businesses that want the system up and running.

Talk to me 📄 CV — AI Agent Systems (PDF)

Agents & Agent Systems

Through this lens

Giants

Salgadelli Atendimento

Vik Nutri

AircTech

Chailter

Organaizer

BBE Tour Generation

Boleiragem

Massinha (Living Interface)

The track record, re-read

Principal Software Engineer · AI Platform Lead

The toolkit for this lens

AI Agent Systems

Multi-LLM Orchestration

Multimodal Pipelines

Want an agent serving your customers tomorrow?