VeritablAI

Alan is "on TV" right after the web socket was connected. He has yet to say anything.
The front page of our website: an overview of the project and a seamless sidebar.

Inspiration

In the modern age, truth is under siege. From AI hallucinations that confidently assert falsehoods to the increasing spread of misinformation in politics, social media, and the news, the sources of untruth are multiplying at an alarming pace. At home, misinformation can mean your loved ones are too afraid to take advantage of healthcare options available to them because of inaccurate information going viral online. Nationally, democracy thrives on the active and honest participation of citizens; misinformation threatens its success by obfuscating or discouraging the best course of action for voters and distorting perceptions of political opponents. Globally, we are seeing a rise in instability; the Oxford Internet Institute found that more than 93% of the countries studied deployed disinformation as part of political communication (https://www.ox.ac.uk/news/2021-01-13-social-media-manipulation-political-actors-industrial-scale-problem-oxford-report).

A lack of access to true and verifiable information is tearing us apart on every level of society, and we cannot address any other issue without the threat of misinformation contaminating the discussion.

There is a way out, and we do not have to put up with this any longer.

VeritablAI was conceived as a response to this crisis. Amid the chaos of misinformation, we seek a path toward clarity: a system that organizes, verifies, and contextualizes verified knowledge at scale in real time. By harnessing reliable data sources and intelligent embeddings, VeritablAI offers a way to cut through the noise; after providing insight on whether the information being told is true or false, it cites its sources, giving users not just access to information, but confidence in its accuracy. In a world increasingly plagued by uncertainty, it is a tool to restore trust in knowledge itself.

Do you want to see what a nation built on truth would look like?

Yeah. Us too.

What it does

VeritablAI processes speech in real time to provide live truth judgments based on well-verified and responsibly-cited sources of true information. We start with Wikipedia--once a bastion for unreliability is now a pillar of well-cited writing.

The process goes like this: A person is talking. Connected to their microphone is a device running VeritablAI, which listens to the speech and records it using realtimeSTT on GitHub. While the speech is transcribed, multiple other processes happen simultaneously: - Vultr hosts a mega corpus of verified information from sources such as encyclopedias. It also hosts a server that supports faster processing over such a large document corpus. - Python functions in tandem with Snowflake's SQL processing power to search the data for the top three to five most relevant paragraphs of information within the corpus. This will be titled "The Evidence." - The Evidence is passed on to a Python function running Gemini API. Gemini handles the Evidence as well as the speaker's Statement and determines if, based on the Evidence, the statement is true. Gemini returns a truthfulness rating. - The rating is output as a signal. - This whole process happens live, running constantly as the user speaks. This avoids the current issue of fact checking coming out after viewers have already moved on--addressing misinformation exactly as it happens will not only prevent viewers from walking away mal informed, but it also encourages speakers to be as honest as possible; if they aren't, there are immediate consequences.

How we built it

We used GitHub to host our code, and Vultr to deploy our server and block storage to contain such a large amount of information. We opted not to use web scraping in this situation, as we felt it was important to control which information is considered "verified" enough in order to prevent our own tool from hallucinating or spreading misinformation further.

Snowflake, SQL, and Python handle the data search algorithm. This then passes to Gemini AI to compare a set of verified true statements with a related statement of uncertain verity.

We are using an HTML template to make our home page for the website https://veritablai.tech, and we have another HTML page to host the television mock-up. While users have their camera and microphone connected, Python code is running in the background to collect audio samples and turn them into speech tokens for use with Vultr.

Challenges we ran into

Our journey was far from straightforward. Vultr account setup was initially impossible because it did not accept our credit card numbers, but fortunately, we were each provided $100 worth of credits. However, it took some persistence for those credits to reach our accounts. Handling large-scale Wikipedia dumps revealed the limitations of common XML parsers and highlighted the need for streaming processing to avoid memory overload. Embedding functions initially failed on long passages, forcing us to rethink how we chunked and structured data. Snowflake connectivity also posed challenges—data might insert successfully, but embedding computations often caused subtle errors that were hard to debug. These technical hurdles were compounded by the inherent complexity of working with massive, unstructured datasets.

Registering the domain name https://veritablai.tech was also a challenge as many failures were involved due to Vultr firewall crashes and the web connection being pointed from Vultr to the tech domain service the whole time instead of the other way around. With enough troubleshooting, we finally registered it successfully.

Accomplishments that we're proud of

Despite the technical and logistical hurdles, we’ve built a fully functioning pipeline capable of processing large volumes of verified information in real time. We successfully integrated streaming transcription, dynamic database querying, and AI-driven truth analysis into a single system. The Evidence now flows seamlessly from Vultr-hosted storage, through Python and Snowflake, into the Gemini API, producing live truthfulness ratings. Beyond the technical achievement, we’ve demonstrated that a tool like VeritablAI can meaningfully empower individuals and organizations to discern fact from fiction as it happens. Mostly, we are proud of creating something that can have a tangible effect on the world as we know it in a way that affects all of us.

What's next for Veritablai

Our next steps focus on expanding both scale and accessibility. We plan to incorporate additional trusted sources beyond Wikipedia, including academic publications and government datasets such as laws. With more time and a bit of funding for a more powerful Vultr server, we could drastically reduce latency further, enabling even faster live feedback. Finally, we aim to broaden access, developing lightweight client integrations so VeritablAI can operate seamlessly across devices and platforms. Our mission is not just to detect falsehoods, but to build a culture of truth: a world where verified knowledge flows freely, transparently, and instantly.

Most of all, VeritablAI is heading toward a nation in need, and with this tool in hand, I believe there is a way for us to find our truth again.