Vinícius – Python, LangGraph, AI agent development

A BI platform for a Brazilian B2B food service marketplace connecting 240+ suppliers with 81 active restaurants and franchises. It supports daily decisions across product, operations, and leadership with real-time order ingestion, KPI dashboards, customer and supplier intelligence, and a Portuguese-language natural language SQL chat for non-technical stakeholders.

• Built and operated a BI platform serving 240+ food suppliers and 81 active customers;
• Improved query performance by 18x on average across 10 query categories, and up to 30x in the best case, using 21 materialized views;
• Shipped a multi-provider AI chat that routes Claude, Gemini, and GPT, with AST-validated NL-to-SQL execution;
• Designed 6-tier RBAC with GDPR retention, 2FA, audit logs, and Upstash Redis edge rate limiting;
• Built N8N workflows for real-time order ingestion and WhatsApp Business API message tracking.

A B2B AI platform for Brazilian internet service providers. It runs WhatsApp-first conversational agents that handle support, sales, billing recovery, and retention end to end, with human handoff only for edge cases.

• Designed and shipped 47 LangGraph agents serving 3 ISPs and handling about 4–5k WhatsApp messages per day;
• Wired 26 tools across ERP, WhatsApp Business API, and the support platform;
• Built LangSmith tracing for 100% of LLM and tool calls, reducing debug time from 30–40 minutes to 5–10 minutes;
• Reached 85% billing recovery on a workload previously handled by 6 human operators;

A WhatsApp AI assistant for a Brazilian B2C e-commerce platform shipping orders nationwide. It resolves most customer questions autonomously across product, billing, and order tracking through a Python RAG pipeline backed by Gemini File Search and a real-time stock tool connected to Bling ERP. The agent runs on N8N with four specialized tools and is validated by 124 end-to-end fixture conversations with a 100% pass rate.

• Built and operated a WhatsApp AI assistant resolving 76% of customer queries autonomously;
• Handled 200–300 daily conversations with a 100% E2E pass rate across 124 fixture conversations;
• Designed a two-layer architecture: static RAG catalog plus real-time stock via a Bling ERP tool;
• Built a hybrid ETL pipeline combining Google Sites scraping, Bagy sitemap validation, and Gemini 2.5 Flash gap-filling;
• Productionized the system on Docker Swarm with an hourly cron, then migrated it to Hostinger for latency and BR data residency.

A WhatsApp AI customer service agent for a regional Brazilian internet service provider. It handles support, billing inquiries, plan questions, and contract management for active subscribers via WhatsApp Business API. The agent queries Voalle ERP in real time to ground responses in customer data, with full LangSmith observability for every LLM decision and tool call.

• Built a LangGraph WhatsApp agent for a regional Brazilian internet service provider processing hundreds of thousands of monthly interactions;
• Drove a 120x improvement in customer wait time, from 4 minutes to 2 seconds, through an N8N to LangGraph migration with feature parity;
• Designed a 7-second Redis-locked message debouncer preventing fragmented replies to burst inputs;
• Engineered a pre-LLM briefing pipeline for cleanup, summarization, and trimming, reducing token cost and hallucination risk;
• Productionized the system on a 2-node Docker Swarm cluster with Traefik, Prometheus, Grafana, and 5-layer resilience.