Victor Goico

Imagine if a single, simple prompt could generate TV-ready video advertisements in 3 or 4 minutes. That's our VideoAI Studio! With optional product/logo reference image upload, we implemented a multi-step video generation pipeline.

Steps include checking for relevant asset images using our RAG pipeline, video and scene composition planning, storyboard image generation, parallel video chunk generation, final stitching, and an optional analysis and music track addition (depending on model used, some support sound out of the box).

Post-processing also allows users to regenerate specific chunks and restitch. The three of us finished this project in one week. I designed and implemented the planning, orchestration, RAG pipelines, and chunk generation.

This project was a great opportunity to combine all of our new skills to this point. Models used include GPT-4o-mini for reasoning steps (checking for relevant assets step, planning step), GPT-4 Turbo for image analysis, Flux Dev for image generation, Flux Dev-ControlNet for image generation with a reference asset, and Google Veo 3.1 OR Minimax Hailuo 2.3 Fast for chunk generation. All deployed using Python FastAPI, Celery, and Redis task queues to scale for multiple videos and users.

SpendSense is another powerful real-world example of how to superpower existing applications with AI. In this case, I built an AI-powered recommendation engine. After being granted permission, the system generates a persona based on a user's spending, saving, and investment habits (using seed data).

The system can then recommend relevant products or learning resources that fit the exact use case. A good saver might want to learn about investing. Someone with many subscriptions might want a subscription manager. The recommendations are tailored to each user's unique financial situation and goals.

Considering the reasoning required for accurate recommendations, I used GPT-4o-mini to power the recommendation engine. Persona assignment uses rule-based logic for the MVP. The solution was deployed to a Python FastAPI back end using OpenAI's SDK.

Remote teams drown in chat messages, and things falling through the cracks is just an accepted fact of life. SOS helps by providing AI-powered summarization, key decision and action-item extraction, and semantic search (keyword weighting and vector embedding comparison). It's a React Native Expo mobile app built with remote workers in mind.

High-priority messages are automatically highlighted in red, ensuring critical communications never get lost in the noise. Real-time chat functionality is powered by Firebase, enabling instant synchronization across the entire team.

I used GPT-4o-mini for all LLM functionality, and OpenAI Text Embedding 3 Small for semantic search capabilities, allowing users to find relevant messages even if they don't remember the exact keywords used.

This was a solo project I created to test my skills with Cursor, AI Agents, Wispr Flow, Firebase, React, and Tailwind CSS. It is a collaborative canvas tool very similar to Figma. I built this as part of the Gauntlet AI fellowship program.

The solution features real-time database sync for user presence, object selection, cursor movement, and object dragging. The rest of the persisted data is handled by the Firebase store, ensuring seamless collaboration between users.

I also used this project to learn how to build my first agent. It's very basic, using the OpenAI SDK and GPT-4 Turbo to assist with object creation and even login forms. The AI assistant can help users create complex shapes and designs with simple, natural-language commands.