Selected Work
RBBT — Sales
Cofounder
A multi-agent AI system that sells through conversation — qualifying customers, building carts and closing orders on WhatsApp and web chat, escalating to a human only when it should.
Multi-Agent Architecture
Four specialized agents (routing, selling, checkout, post-sale) each pinned to a least-privilege tool subset, so a failure in one never widens the blast radius of another
Platform-Agnostic Core
A brand-agnostic engineering layer with pluggable commerce adapters: the same selling logic runs on Shopify and Nuvemshop without touching core
Trust and Safety
Prompt-injection detection, tool-call anomaly gating and per-brand rate limits, all fail-closed to human escalation
Evaluation as Infrastructure
Agents are scored continuously on routing accuracy, recommendation quality and checkout conversion; autonomy is earned against metrics, not assumed
Typescript - Anthropic Claude - Shopify - Nuvemshop - WhatsApp Cloud API - PostgreSQL
view project →Fomenta
Cofounder & CTO - Programa Nascer - FAPESC
An end-to-end SaaS that finds public research funding for Brazilian researchers and matches it against their profile. I own the architecture from the ingestion pipeline to the payment layer.
Event-Driven Ingestion
Designed a nine-Lambda pipeline fanned out over SQS that scrapes federal and state funding agencies, then runs a three-stage LLM chain to parse, validate and normalize every opportunity
Failure Containment
Dead-letter queues at each stage, deduplication before extraction, and CloudWatch alarms on queue age and DLQ depth — the pipeline degrades loudly instead of dropping data silently
Semantic Matching Service
A separate async service ranks opportunities against a researcher's profile, cached in DynamoDB under a key stamped with the prompt version, so improving the prompt invalidates stale results on its own
Commercial Layer
Role-based access, Stripe subscriptions and payment webhooks, with per-tier usage quotas that keep LLM cost bounded per user
Next.js - Typescript - AWS Lambda - SQS - DynamoDB - Supabase - LangChain - Puppeteer - Stripe
view project →SwellGuidePaused
A daily mailing service that turned raw weather data into plain-language surf forecasts.
Product Growth
Designed, launched and grew the service to 50+ daily active users receiving automated email reports
Decoupled Architecture
Built a modular system that allowed new surf spots to be added without significant code changes, making it cheap to scale coverage
Automation Workflow
Created an automated Node.js workflow orchestrated with LangChain to consume weather APIs, process insights via LLM, and dispatch campaigns
Node.js - LangChain - API Integrations - Tailwind CSS - Javascript - AWS