João Vitor Chaves de Souza

Building V4H, Wisecare's telehealth platform used by SUS, Brazil's public health system, integrated with the AGHUse health record, CNES, the national registry of health facilities, and Keycloak authentication. I built the processing and persistence of consultation transcripts and an AI layer on top of them: citation-grounded search, clinical fact extraction per participant and SOAP progress-note drafts. I work on the Node and TypeScript backend (Express, TypeORM, InversifyJS, BullMQ and RabbitMQ, Socket.IO) and on the React frontend with Redux-Saga, and I am rebuilding session and video-call management to replace the Jitsi dependency with an in-house solution.

Research, development and innovation with industry partners, focused on LLM-based agents, OCR and computer vision. Projects: Editora BP (May 2026 — present), an AI platform for producing annotated exams, with a 10-step LangGraph workflow, RAG over curated sources and expert review, which cut the first draft of an exam from one or two days to about ten minutes and the cost per question by 10x; and TutorIA (Dec 2025 — Mar 2026), ENEM tutoring on WhatsApp that solves questions from photos, grades essays and answers doubts with RAG.

At the Technology and Artificial Intelligence League, the first academic AI league in Paraíba, I take part in the Computer Vision/Sports directorate. With the team, I develop an automated scouting system for 3x3 basketball that extracts player metrics from broadcast footage, adapting a 5x5 pipeline (detection, segmentation, court keypoints and pose) to half-court play and evaluating how well the models generalize. In progress.

With the trainee cohort, I built Pauta, a civic-tech platform where citizens report urban problems with a photo and location. A vision LLM classifies each report (with manual review when confidence is low), reports become geolocated points in PostgreSQL/PostGIS shown on a map and community feed, and notifications run on an event outbox consumed by Celery workers.

Owned the automated ingestion pipeline for LOTEP, the Paraíba state lottery: it detects new spreadsheets from every licensed operator in the state, matches each one to a known layout or infers and registers a new layout automatically, and runs vectorized auditing and cleaning up to the load into a PostgreSQL data warehouse, covering over 20 GB of data and millions of betting records. Metadata-driven JSON configuration, anti-fraud validations, event-driven architecture on Kafka and a FastAPI gateway validated by schema, LGPD-compliant. Cut data preparation for the analytics team from 6 hours to 40 minutes.

Agent generator that solved a bottleneck in CrewAI: instead of writing code, the user defines the agent's role, goal and context through an interactive interface, and the application generates ready-to-use code and configuration.

Free programming, projects and mentoring program at UFPB, run by students, with classes in Python, web and data.

First place among six teams in the program hackathon (May–Jun 2025), with PixelMind.

LembreMe

Personal assistant that runs entirely on WhatsApp: finances, tasks, reminders, goals, and routines in natural language. Lead author, architecture and development.

Python · FastAPI · Node.js · whatsapp-web.js · LangChain · GPT-4o-mini

Pauta

Municipal political transparency: citizens report an urban problem, track its resolution, and get closer to the people who represent them.

Python 3.12 · FastAPI · SQLAlchemy · Celery · Next.js 16 · React 19

PixelMind

Video editor that accepts natural language commands, first place at the Trilha hackathon. Backend architect.

Next.js · TypeScript · Tailwind CSS · Node.js · Python · FastAPI

Vero

A clinical platform that follows the consultation in real time and runs the patient's pre-visit intake on WhatsApp. Lead author, clinical LLM pipeline, pre-visit intake agent, and document import.

TypeScript · React 19 · Vite · Tailwind CSS 4 · TanStack Query · Fastify