Founding AI Full-Stack Engineer
🌐 Portfolio · LinkedIn · GitHub · Medium
🥉 3rd Place · ADAPTA AI Challenge Hackathon
São Paulo · August 2025 · R$ 100.000 prize · Top 3 of 6,000+ startups
🌐 adaptamidia.com
AdaptaMidia is an AI-powered platform for Out-of-Home (OOH) advertising — combining Computer Vision and LLMs to turn traditional outdoor media into a measurable, data-driven channel.
What it does
| Capability | Description |
|---|---|
| 👁️ Real-time audience analytics | computer vision detects quantity, age range, gender and emotional response (LGPD-compliant, privacy-first) |
| 📊 Measurable ROI | campaign analytics and performance metrics instead of estimates |
| 🤖 Automatic campaign optimization | AI adjusts creative by location, time window and audience profile |
| 🗺️ Actionable insights | peak hours, attention maps, location-specific performance |
Who it serves
| Audience | Value |
|---|---|
| 🏢 Establishments (malls, elevators, restaurants) | monetize spaces with 30–50% revenue share |
| 📣 Advertisers | data-driven campaign management with real-time optimization |
| 💼 Investors | validated tech in an R$ 8B Brazilian market |
My role — built solo, end-to-end
I designed, built and shipped the entire product by myself:
| Area | Contribution |
|---|---|
| 🎨 Frontend | React + TypeScript dashboard and display surfaces |
| ⚙️ Backend | Node.js / NestJS APIs and services |
| 🧠 AI / Computer Vision | detection, demographic inference and attention tracking |
| ☁️ Infrastructure | deployed on Google Cloud with CI/CD |
| 🗄️ Data | MongoDB powering insights and dashboards |
🏗️ Architecture
Cloud-native, end-to-end system I designed and built on Google Cloud:
| Layer | Implementation |
|---|---|
| 📱 Clients | a Flutter / Android app runs the YOLO (Ultralytics) computer-vision model on-device (edge inference, privacy-first), and a React admin dashboard is served from Vercel. |
| 🛡️ Edge | all client traffic flows through Cloudflare (DNS, CDN, TLS, protection) before reaching the backend. |
| ⚙️ API | a containerized backend on Google Cloud Run (serverless, auto-scaling) orchestrates the whole platform. |
| 🔥 Firebase | handles auth and realtime sync between the mobile app and the backend. |
| 🧠 AI | Google Vertex AI powers the managed ML layer for demographic inference and audience analytics. |
| 🗄️ Data | MongoDB stores insights and campaign data; Redis serves as cache and job queue; Google Cloud Storage holds media assets and processed outputs. |
Sponsorships & Credits
| Sponsor | Support |
|---|---|
| ☁️ Google Cloud | US$ 2,000 in cloud credits |
| 🍃 MongoDB for Startups | database credits & startup benefits |
| 🟢 NVIDIA Inception (for Startups) | GPU credits & partner program access |
Traction
| Milestone | Details |
|---|---|
| 🏆 Finalist at Adapta Summit 2025 | top 3 out of 6,000+ startups |
| 🤝 Strategic partnerships | (BitCoffee + technology providers) |
| 📺 Installations | Running installations in premium commercial elevators |
| ⚡ Enterprise-grade infrastructure | 99.9% uptime |
🎬 Videos
1. Me testing the AI — real-time detection
That's me in front of the camera: the model picks me up live, tracks attention and reads emotional response in real time.
demo.mp4
2. In-elevator deployment — counting & demographics
AdaptaMidia running inside a commercial elevator: detecting and counting people, and inferring gender and average age range per passenger group.
case_1.mp4
3. Detection model — bounding boxes in action
Example of the computer vision pipeline: the AI draws bounding boxes around each person detected by the underlying CV models.
feature_people_detected.mp4
4. Dashboard — generated insights
Insights being produced from the captured data: audience breakdowns, peak hours and attention maps feeding the analytics dashboard.
feature_insights.mp4
Natiely is a fantasy hack-and-slash built around a heroine's journey through a labyrinth to rescue her brother. It began as a Three.js prototype and evolved into a native Unity 3D game, with a cinematic website, a player portal and a Go backend.
| Area | Implementation |
|---|---|
| 🎮 Game | Unity 6 / C#, Universal Render Pipeline (URP), Cinemachine, Input System and Addressables for rendering, cameras, controls and game content. |
| 📱 Platforms | Native Android, iOS and macOS builds, with Google Play and Apple App Store integrations. |
| 🌐 Website & player portal | React, TypeScript, Vite, TanStack Router / Query and Tailwind CSS for the cinematic landing page, player dashboard and web store. |
| 🧊 3D workflow | Meshy for AI-generated 3D assets, Blender + MCP for modeling, rigging and skin-weight adjustments, and Mixamo for character animations, with mesh and texture optimization for Unity. |
| ⚙️ Backend & accounts | Go / Echo, Firebase Authentication, MongoDB and Redis for player accounts, game services, persistence and caching. |
| 💳 Commerce | RevenueCat for native purchases through Apple App Store and Google Play, Stripe Checkout for the web store, and a Go backend for Gold balances, purchase fulfillment and entitlement verification. |
| 📊 Monitoring & engagement | Sentry for error monitoring, PostHog for web analytics and OneSignal for push notifications. |
🎬 Game demo
demo.mp4
🎬 Landing page
landing-page.mp4
🎵 Original soundtrack
ClassGame is the optional homework that takes notes for the teacher. Built for the All Things Agentic Hackathon in the Collaborative Partner — in depth category, it turns a teacher's lesson and worksheets into short, personalized 3D review missions.
🎮 Student game · 🧑🏫 Teacher dashboard · 🎬 Full demo · 💻 Source code
🎬 Product demo
Mia turns a teacher's lesson into a mission, guides the student through the 3D classroom and returns evidence to the realtime dashboard:
What it does
| Capability | Description |
|---|---|
| ✨ Generates grounded missions | Mia, the Multimodal Intelligence Agent powered by Gemini on Vertex AI, transforms the teacher's prompt and attachments into questions, misconception traps and traceable knowledge chunks. |
| 🧠 Adapts every turn | stateful, multi-turn tutoring retrieves lesson context, attempt history and student memory to vary explanations, difficulty and next actions. |
| 🎮 Makes review playable | students complete ten-minute missions in a responsive Three.js classroom and resume at the exact saved question and phase. |
| 📊 Closes the loop for teachers | the realtime dashboard shows progress, RAG provenance, learning evidence, recurring class patterns and a prompt for the next lesson. |
| 🛡️ Keeps student and teacher contexts separate | server-side grounding, student-safe projections, leak guards and deterministic AI fallbacks protect private observations. |
What I built
| Area | Contribution |
|---|---|
| 🖥️ Web experience | Next.js 16, React 19, TypeScript, Tailwind CSS and Three.js for the student game and teacher dashboard. |
| ⚙️ Agent backend | a Go/Echo API orchestrating mission generation, grounded tutor turns, learning traces and class-level synthesis. |
| 🤖 AI and memory | Gemini through Vertex AI, structured outputs, realtime RAG, persistent student memory and evidence-backed teacher observations. |
| ☁️ Production platform | two Cloud Run services backed by Firestore, Cloud Storage, Cloud Tasks, Redis, Secret Manager, Cloud Build, Artifact Registry and Cloud Logging. |
🏗️ Architecture
Production architecture for the complete teacher → Mia → student learning loop:
| Flow | Implementation |
|---|---|
| 🧑🏫 Teacher flow | the dashboard sends lesson context to the Go API, which uses Gemini to build versioned missions and publishes student-safe content through Firestore. |
| 🎮 Student flow | the 3D game combines realtime mission state with API-authorized tutor turns, exact checkpoints and persistent learning memory. |
| 🧠 Agent loop | each turn retrieves relevant knowledge chunks and prior evidence before Gemini chooses the explanation and next action. |
| 📊 Learning intelligence | Cloud Tasks finalizes attempt traces and class synthesis asynchronously, then Firestore streams the results back to the teacher. |
Maze is a 3D maze game with procedural worlds, collectible gems and enemy encounters — playable locally, in ranked solo runs or online groups.
| Area | Implementation |
|---|---|
| 🖥️ Web & gameplay | Next.js, React and TypeScript, with Three.js for 3D rendering and rot.js for procedural maze generation and pathfinding. |
| 🧊 3D workflow | Blender + MCP (Model Context Protocol) for working on 3D character models, with optimized GLB models and WebP textures for the browser. |
| 🌐 Online | Firebase Authentication, WebSocket multiplayer and leaderboards. |
🎬 Gameplay
🧊 3D character workflow in Blender
blender-modeling.mp4
As Founding AI Full-Stack Engineer, I'm building CoGrader from zero — an AI-powered grading assistant that helps teachers grade student work faster and more consistently, aligned to their own rubrics.
What I build
| Area | Contribution |
|---|---|
| 🤖 AI features | LLM-powered grading, feedback generation and rubric alignment, with Fine-Tuning and MCP (Model Context Protocol) |
| 🧩 LMS browser extension | brings CoGrader into the tools schools already use (Canvas, Schoology, Blackboard), cutting onboarding friction and driving product adoption |
| 🖥️ Product | React + Mantine front-end, Node.js / Firebase backend, PostHog analytics |
| 🌐 Cross-browser | Plasmo-based builds for Chrome, Firefox and Safari |
🧩 LMS Extension — demo
The browser extension guiding a teacher through setup right inside their LMS:
🔗 Try it on the Chrome Web Store
GoatVPN is a NordVPN-style consumer VPN built on WireGuard — fast, private and unlimited access across a worldwide server fleet (live on Google Play). As CTO, I architected and shipped the entire platform end-to-end and built and led the engineering team.
🎬 React Native mobile demos
Screenrecorder-2025-03-10-08-44-07-271.mp4 |
Screenrecorder-2025-03-13-16-16-24-444.mp4 |
Screenrecorder-2024-12-13-00-54-01-33.mp4 |
What it does
| Capability | Description |
|---|---|
| 🌍 Servers worldwide | connect through a global WireGuard fleet for low-latency, secure tunnels |
| ⚡ Blazing-fast & unlimited | modern WireGuard protocol, no bandwidth caps |
| 🔒 Privacy-first | fully encrypted traffic with a no-logs posture |
| 📺 Unblock content | access geo-restricted streaming and sites |
| 📱 Multi-platform | React Native mobile app + web client |
My role — architected, shipped & led the team
| Area | Contribution |
|---|---|
| 📱 Clients | React Native mobile app and a React web client (auth, server picker, subscription flow), each receiving per-user WireGuard configs from the API |
| 🧩 Mobile stack | React Native 0.76 / React 18 with React Native Paper, Zustand, React Router Native, Reanimated, Gesture Handler, Maps, Skia, Victory Native, Pusher, AsyncStorage, permissions and device-info integrations |
| 🔐 VPN infrastructure | WireGuard core with a large server fleet orchestrated via WGDashboard, provisioned across DataPacket and IS Hosting for global, low-latency coverage |
| ⚙️ Backend | NestJS API on AWS issuing configs and managing the fleet |
| 🗄️ Data | MongoDB for users, subscriptions and server metadata; Redis for queue & cache |
| 💳 Billing | multi-gateway across Stripe (cards), Cryptomus and NoxPay (crypto / alternative checkout) |
| 📈 Growth & ops | ClickUp for delivery, Slack for messaging & alerts, and Meta / Google tag managers wired to subscription events for conversion tracking |
Engineering highlights
| Highlight | Details |
|---|---|
| 🛰️ Fleet management | Automated provisioning and management of a multi-provider server fleet through WGDashboard |
| 💰 Resilient multi-gateway billing with crypto support | a single provider outage never blocks conversions |
| 🏗️ Cloud-agnostic split | AWS for the API, independent bare-metal/VPS providers for VPN egress |
🏗️ Architecture
Production architecture for the VPN platform:
| Layer | Implementation |
|---|---|
| 📱 Clients | iOS, Android, React web and a Plasmo browser extension consume the same API surface. |
| ⚙️ API | a NestJS backend on AWS centralizes auth, subscriptions, server metadata and VPN provisioning flows. |
| 🗄️ Data | MongoDB Atlas stores users, subscriptions and server metadata; Redis handles cache and queue workloads. |
| 💳 Payments | Stripe, Cryptomus and NoxPay support card, crypto and alternative checkout flows. |
| 📈 Growth & monitoring | subscription events feed Google Tag Manager and Meta, while Slack and New Relic support ops and observability. |
| 🔐 VPN fleet | WireGuard servers are managed through WGDashboard and provisioned across DataPacket and IS Hosting. |
WPP AI is a multi-tenant NestJS API that connects WhatsApp accounts to LLM-powered support triage. It ingests messages through Baileys, records conversation history, analyzes support requests with LLM providers, and exposes conversations plus AI audit trails to an admin frontend.
What it does
| Capability | Description |
|---|---|
| 💬 WhatsApp ingestion | connects to WhatsApp through Baileys, maps incoming private/group messages and persists normalized conversation history. |
| 🏢 Multi-tenant routing | resolves tenant, WhatsApp account, environment and bot profile before persistence or AI execution. |
| 🧠 AI support triage | uses deterministic reviewers first, then queues LLM analysis with guardrails, priority, suggested action, audience and evidence. |
| 🔁 Provider fallback | runs Gemini first and falls back to OpenAI, with model fallback per provider. |
| 🧾 Auditability | stores request text, prompt, analyzed messages, provider/model, token usage, guardrail result and error details for every analysis. |
| 🔐 Admin API | exposes authenticated conversation and analysis endpoints for a frontend dashboard. |
🏗️ Architecture
| Layer | Implementation |
|---|---|
| 📲 WhatsApp adapter | Baileys receives messages, maps protocol payloads, loads media when possible and sends quoted replies. |
| 🧩 Orchestration | the orchestrator decides between direct deterministic replies and queued LLM jobs. |
| 🧱 Persistence | MongoDB stores tenants, accounts, bot profiles, conversations, messages and analysis snapshots. |
| 🚦 Queue | Redis/Bull decouples WhatsApp ingestion from model execution, retries and final replies. |
| 🛡️ Guardrails | prompts and model outputs are wrapped and validated before replies reach users. |
| 🤖 AI providers | Gemini and OpenAI are abstracted behind provider fallback logic for reliability and cost control. |
🎬 Demo
WPP AI handling a WhatsApp support conversation through the LLM orchestration flow:
demo.mp4
Velrino UI is a library of accessible, themeable React components with optional sci-fi surfaces, built for independent React applications. It grew out of HackUniverse's gallery, dialog frame and globe loader, while keeping each application's colors, fonts, brand, localization and business logic independent.
🌐 Interactive gallery · 💻 Source code
What it provides
| Area | Description |
|---|---|
| 🧩 React components | Buttons, inputs, selects, dialogs, panels, tabs, sheets, step navigation and loading feedback. |
| ♿ Accessible interactions | Keyboard navigation, focus management, localized controls and reduced-motion support. |
| 🎨 Application-owned themes | Semantic CSS tokens preserve each application's palette, fonts and branding. |
| ✨ Optional sci-fi surfaces | Standard surfaces by default, with decorative frames enabled per component or through a shared provider. |
| 🌐 Interactive gallery | Live component examples with English/Portuguese copy, theme switching and surface previews. |
| 🤝 Shared across projects | Used by my portfolio and HackUniverse. |
I'm passionate about pushing the boundaries of what's possible with AI. Currently building AI-powered features at CoGrader using LLMs, Fine-Tuning, and MCP (Model Context Protocol).
What I work with:
LLM · Fine-Tuning · Model Context Protocol (MCP) · OpenAI APIs (ChatGPT, DALL-E, Whisper) · Computer Vision · OCR · Leonardo AI API · Prompt Engineering
Full-Stack Engineer with 12+ years of experience (since 2014) building scalable products from scratch. I've gone from intern to CTO, shipping products used by millions of users across banking, fintech, VPN, and edtech.
Currently: Founding AI Full-Stack Engineer at CoGrader — building the product from zero with Node.js, Firebase, React & AI.
Previously:
- CTO / Tech Lead at GoatVPN — built the entire platform and hired the engineering team
- Tech Lead at Itaú BBA — led authentication & security squad at Latin America's largest investment bank
- Tech Lead at GO-K Digital (Grafeno Digital) — led the Credit Emission fintech project
- FIC Engineer at Trustly — payment gateway integrations across North America & Europe
- Software Engineer at Itaú Unibanco — mobile banking apps for millions of users
- Software Engineer at Accenture — enterprise solutions for Porto Seguro
Languages
Backend
Frontend
Mobile
Databases & ORMs
Cloud & DevOps
Messaging & Integrations
Also: Stripe · Pusher · Twilio (SMS/WhatsApp) · SendGrid · Slack API · Cloudinary
Monitoring & Observability
Also: New Relic · DataDog
Testing
Tools
Original background videos used by velrino.co, served directly from this repository.
Desktop — landscape
Mobile — portrait
- ChatGPT in WhatsApp — Integrating ChatGPT into WhatsApp using Facebook's official API with NestJS
💬 Let's build something together
📧 velrino@gmail.com ·
LinkedIn ·
🌐 Portfolio