VideoSignal

Product: VideoSignal
Year: 2025
Project Type: AI media analysis platform and API
Stack: Next.js, Supabase, AWS S3, Stripe, Google Gemini, OpenAI, Whisper

Project Overview

VideoSignal takes a piece of media and tells you what is inside it. You give it a video, an audio file, an image or a PDF, and it gives back a description, a full transcript, the tone and the mood.

I worked across the product. I built how people sign in, how developers get access through API keys, how usage is paid for with Stripe credits, and the AI analysis that sits at the heart of it.

What I Built

Sign in and accounts

• Email and password accounts on Supabase Auth
• Sign in with Google
• Protected pages, so only signed in users reach the app

API keys

• Keys that let developers send media to VideoSignal from their own apps
• A place in the dashboard to create and manage them

Billing with Stripe

• A credits model: start free, pay as you go, or move to a monthly plan
• Credits used as media is analyzed
• Checkout and plans handled through Stripe

AI analysis

• Uploaded media stored on AWS S3
• Descriptions, transcripts, tone and mood for video, audio, images and PDFs
• Google Gemini and OpenAI for analysis, and Whisper for transcription

Tech Stack

• Next.js: The web app and its API
• Supabase: Accounts, sign in and data
• AWS S3: Storage for uploaded media
• Stripe: Credits, plans and checkout
• Google Gemini and OpenAI: Understanding what is inside each file
• Whisper: Turning speech into text

The Hardest Part: Audio

Video, images and PDFs were hard in their own ways, but audio was the one that really pushed me. A recording is not always someone talking. Very often it is a song.

A transcript alone is no help for music. If someone uploads a hip hop track, the useful answer is that it is hip hop, what the mood feels like and what the song is about, not just a wall of lyrics. Many files are a mix of both, like a podcast with music in the background or a video with a soundtrack under the voice.

I spent a lot of time on this. The analysis had to work out whether it was hearing speech, music or both, and then describe each in the right way. For music that meant recognizing the genre, hip hop, pop, rock and so on, along with its tone and mood. For speech it meant a clean transcript and a sense of how the speaker sounded.

Getting this right is what made the audio results feel useful instead of generic.

Key Learnings

• Audio is never just speech, and the analysis has to know what it is listening to
• Sign in, keys and billing are the parts users trust you with, so they have to be solid
• A simple credits model is easier for people to understand than complex plans
• AI results are only as good as the questions you ask of the media

Conclusion

VideoSignal let me work on a product from the front door, signing in and paying, all the way through to the AI that does the real work.

The audio work taught me the most. Teaching software to tell a hip hop track from a conversation turned out to be a much bigger problem than it first looks.