satoricheck.com

Satori is a light weight fact checker with multi-agent setup; verify claims in live audio / text and pitchdecks instantly, because we need more fact-based conversations.

Comment

Inspiration

People share some garbage quality info online. Wrong facts, misleading information, ai hallucinations that all gets reiterated as truth. But questioning every multi-sentence statement within a politicians speech and in news is both hard and mentally tiring. "Why not build a real-time fact-checker myself?", I thought, while playing around with the gemini ai vibecode tool. I hastily sketched the idea, procrastinated a while watching youtube videos, then I started vibe coding the hollow ui ... the next time I looked up from my screen 7 hours had passed.

What satori does

It transcribes audio, identifies nested or complex claims within paragraphs, and verifies them against primary sources. It's a multimodal input processing (Text/Audio transcription) and contextual factchecking (sources, social) tool, sparkled with a Document/Presentation (PDF/Pitch deck) auditing module. And because people are annoyed when being called out, it has a social share function. Naturally.

How I built it

I am not a developer. A big part was using gemini as technical tutor. It led me through the entire implementation, explained functions and modules, created UI mockups and more.

Building The Google Antigravity IDE blew my mind. I used mainly gemini 3 pro & 3 flash in tandem with claude opus / sonnet for all coding tasks, back and forth with my other gemini chats. I let the agent write and run all tests letting it use the chrome integration with a local host for the frontend tests, plus all e2e testing sessions and so on.

Deploying I used the google cloud for everything SaaS, with Cloud Run, Cloud SQL, 0Auth. I let gemini research the tools and then explain to the agents how to use them, how to test and what to consider about and then the agents run the entire plan and implement process themselves.

App logic and reasoning Gemini does the logical branching, it contextualises and identifies claims both in text and pitchdecks (using the thought signature and context caches) It is the primary source fact checker, thus, it decides whether it wants to pass around the claims to the other agents for meta checks/social searches (grok), if it wants to do a check directly, send it as batches and so on, and then it brings it all back together.

Challenges I ran into

haha, please 😂 Learning the operational parts such as API setups, Google Cloud tools setup and deploying via github and so on from scratch took a while. I also did not understand how to integrate the thought signatures and context cachings efficiently, so what I did was building what I wanted with hardcoded rules (regex etc), and only when the logic was sound, we started replacing what the app does with increased agentic liberty, optimising for costs and api usage. I also tried bugfixing for days, until I learned the dev console in chrome has an ai helper that explains errors I can use as feedback haha.

Accomplishments that I'm proud of

I went all the way with an ai as tutor, from zero to a live, micro-SaaS with actual paying users within a few weeks, with a scalable architecture using a multi-agent setup and real-time API integrations, while working full time. 100% Madman Pride regardless what anyone says.

What I learned

Everything operational. From scratch. Code architecture, security protocols, API orchestration, database management, and how to prompt-engineer agents for coding tasks. Pretty awesome.

What's next for satoricheck.com

The goal is so it can be left running to analyse conversations / watch videos with you, continuously checking for claims while they are stated. It can decide by itself which agents are needed to optimise for cost efficiency.

  • First, Video Fact-Checking, I want to use the extended context window of gemini 3 pro more, and improve the live video / audio factchecking through this.

  • Second, Robustness and logic, I want to make it able to run 24/7 but this needs better use of the context window I have planned for, but I am aware that I am getting to the point where I need some professional insights.

  • Third, Building a Satori extension for Chrome, Opera/Brave., for a true 24/7 checker directly in browser.

I want to be watching a live stream, while Satori real-time audits the speaker in the background. !

Built With

Share this project:

Updates

posted an update —

Last update: I added a media authenticity checking module for pics and videos using the new gemini functions to check for ai signs. will build this out to deepfake checking down the line but probably train my own model to specialise a bit on that. Overhauled some stuff with thougt signatures and context cache as well, so I could reduce API calls and load.

Decided to publish the current repository to my github as public. Will be the last time I write an update here. Cheers

Log in or sign up for Devpost to join the conversation.

posted an update —

Just saying hi, I am thinking about starting with the video module next.

Not sure if I should go for fact checking a live stream or batch checking a video like youtube first, and how to make this usable without blasting the API with hundreds of calls per user per minute.

Log in or sign up for Devpost to join the conversation.

posted an update —

Found several issues with the agentic decision making and context caching that were not properly implemented when I submitted the code. Have updated the live version today (not the submitted code-repo) with an improved agentic triage, and then also redid the batching logic to reduce context-loss. Quite a significant improvement in speed with the micro batching as well.

Log in or sign up for Devpost to join the conversation.