Complex problems.
Thoughtful software.

COMPUTER VISION / AI

Adapta Mídia

Turn street activity into useful advertising insights.

• Flutter
• React
• YOLO
• Cloudflare
• Cloud Run
• Firebase
• MongoDB
• Redis
• Vertex AI
• Cloud Storage

Award-winning platform for outdoor advertising analysis and optimization using Computer Vision and Artificial Intelligence

An Android app built with Flutter runs YOLO (Ultralytics) for on-device computer vision, while the admin panel is a React web app deployed on Vercel. Both consume the same backend API.

Cloudflare sits in front of the platform as the edge layer — handling DNS, CDN caching, TLS, and protection — and routes every client request to the backend API.

The API runs on Google Cloud Run (serverless containers, auto-scaling). Firebase handles auth and realtime needs, MongoDB is the primary database, and Redis powers caching and the job queue.

Google Vertex AI provides the managed machine-learning layer for advertising analysis, and Google Cloud Storage holds media assets, uploads, and processed outputs.

GAMES / INDEPENDENT

Natiely

A world built from a lifelong love of games.

• Unity
• C#
• Game development

My independent fantasy game, bringing together game development, art, music, and the experience of building a product from the ground up.

EDUCATION / 3D

ClassGame

A 3D classroom where AI helps students learn.

• Next.js
• React
• TypeScript
• Tailwind
• Three.js
• Go
• Echo
• Gemini
• Vertex AI
• Firestore
• Redis
• Cloud Tasks
• Cloud Run
• PostHog

ClassGame turns a teacher's lesson and worksheets into a short 3D review mission. Mia asks why a student chose a wrong answer, retrieves grounded lesson context, adapts the explanation, remembers how that student learns across sessions, and sends observable evidence back to the teacher in real time.

Next.js 16, React 19, TypeScript, Tailwind, and Three.js deliver two connected experiences. Students play adaptive missions in a mobile-friendly 3D classroom; teachers create, revise, publish, and monitor missions from a live evidence dashboard.

A Go and Echo API re-verifies Firebase identity, enforces class and mission authorization, pins every attempt to an immutable published revision, grounds question data server-side, and orchestrates synchronous tutoring plus asynchronous learning work.

Mission, Mia Tutor, and Learning agents call Gemini through Vertex AI with structured output. A policy and grounding gate combines retrieved lesson chunks, recent observable events, tutor history, persistent student memory, and class patterns before each model turn.

Firestore stores classes, versioned missions, exact attempt checkpoints, tutor turns, learning traces, long-lived student memory, and class synthesis. Redis enforces distributed quotas and health gates, while Cloud Storage keeps canonical published attachments.

Cloud Tasks delivers OIDC-authenticated jobs for attachment copies, learning-trace finalization, and the abandonment watchdog. Cloud Build and Artifact Registry deploy both services, while Cloud Logging and PostHog capture operational and product signals.

Browser Firestore access is read-only and role-scoped; students receive only safe projections. Leak guards protect teacher-only context, event IDs make tutor turns idempotent, quotas bound model spend, and every agent contract has a deterministic fallback.

MESSAGING / AI

WPP AI

AI conversations, inside WhatsApp.

• Node.js
• NestJS
• React
• Baileys
• MongoDB
• Redis
• Bull
• Gemini
• OpenAI

Connecting WhatsApp with AI through a backend built around messaging, integrations, and reliable processing.

WhatsApp users and groups message the bot, while the React admin panel reads conversations, analyses, and authentication state through the same API.

Every message is resolved to a tenant, WhatsApp account, environment, and bot profile before persistence or AI work. That keeps customers, numbers, prompts, and histories isolated inside one deployment.

The NestJS API owns the Baileys adapter, auth, conversation, and orchestration modules. MongoDB stores tenants, accounts, conversations, messages, and analyses; Redis powers the Bull queue and operational cache.

Deterministic reviewers answer cheap simple cases without an LLM call. When a model is needed, the system wraps prompts with guardrails, uses Gemini as the primary provider, and falls back to OpenAI.

WhatsApp ingestion records the message and enqueues a job, keeping the socket responsive. A worker processes the analysis, validates output, completes or fails the audit snapshot, and sends the answer through the WhatsApp sender.

PRIVACY / INFRASTRUCTURE

GoatVPN

A VPN with the infrastructure to back it up.

• WireGuard
• WGDashboard
• NestJS
• React
• AWS
• MongoDB
• Redis
• Stripe
• Cryptomus
• NoxPay

A NordVPN-style consumer VPN powered by WireGuard, with a fleet of servers managed through WGDashboard, a NestJS backend, a React app, and multi-gateway billing across Stripe, Cryptomus, and NoxPay.

An Android app and a React web client let users browse server locations and connect. Both authenticate against the same backend API and receive their WireGuard configuration from it.

The product runs on WireGuard, with a large fleet of servers orchestrated through WGDashboard. Capacity is provisioned across DataPacket and IS Hosting infrastructure for global, low-latency coverage.

A NestJS backend running on AWS exposes the API. MongoDB stores users, subscriptions, and server metadata, while Redis powers the job queue and caching layer.

Billing runs across multiple gateways — Stripe for cards, plus Cryptomus and NoxPay for crypto and alternative checkout — so users can subscribe through whichever method fits them.

ClickUp drives task management and Slack handles team messaging and system notifications. Subscription events are forwarded to Meta and Google tag managers to tag users and track conversions.

EDUCATION / AI

CoGrader / Cortex

AI feedback that reflects each teacher’s grading style

• React
• TypeScript
• Node.js
• Go
• Redis
• BullMQ
• Vertex AI
• Gemini
• JEV
• Firestore

I helped build Cortex, CoGrader’s new AI system, to adapt grading and feedback to each teacher as the product grew. My work spans handwritten document processing, reliable job queues, AI orchestration and the browser extension.

As CoGrader grew, one common AI approach could not reflect every teacher’s expectations. The challenge was to make feedback more personal while reliably processing the different files teachers upload.

Processed PDFs, scanned pages and handwritten student work with Go and Node.js. Redis-backed queues and idempotent processing let jobs retry without repeating their effects.

Orchestrated small agents with focused tasks and used JEV for classification, avoiding LLM calls for decisions that did not need them. Gemini 3.7 runs through Google Cloud Vertex AI, supported by credits received from Google.

Also contributed to the notification system, backend services and the React application, supporting the experience beyond the AI response.

BROWSER EXTENSION / EDUCATION

CoGrader Extension

CoGrader, where teachers already work.

• React
• TypeScript
• Plasmo
• Tailwind
• EdLink

Built the React and TypeScript browser extension from scratch, bringing CoGrader into Canvas, Blackboard, Schoology, Brightspace, Microsoft Teams for Education and Google Classroom. It increased interest in the main product and improved conversion by removing integration friction.

React and TypeScript power the interface, with Plasmo organizing the browser extension. LMS integrations bring grading into the tools teachers already use, helping them discover the main CoGrader product without changing their workflow.