VerifyDFW

Inspiration

During Winter Storm Uri in February 2021, over 200 Texans died while millions lost power for days. As that crisis unfolded, neighborhood WhatsApp channels, Nextdoor threads, and family group chats became the one of the primary sources of safety updates for residents trying to figure out what was happening. Winter Storm Uri, and many of the crises that followed it, showed us that when severe weather hits the DFW area, information moves incredibly fast, but the truth clearly struggles to keep up. We saw this dynamic play out firsthand in our own community networks.

By the time an official municipal agency or news outlet could post a verified update, many different versions of the situation had already spread across our neighborhoods. Sometimes people were panicking over rumors of permanent grid failures, whilst others ignored actual localized hazards.

The people dealing with the brunt of this chaos are local volunteers, more specifically known as block leaders or community coordinators. Cities across Texas formally rely on networks of neighborhood block leaders, HOA coordinators, and community organizers because trusted local figures can text information to residents much faster than a state agency can broadcast it. But during actual emergencies, these volunteers are often left totally empty handed. They are forced to act as emergency dispatchers, making high stakes judgment calls about what safety warnings to forward to hundreds of families using nothing but their gut, or worse, pieces of information spread online. We built VerifyDFW to give them a tool so they can stop guessing in the dark, aiming to help them, and as a result, dozens of families relying on them.

What it does

VerifyDFW is a tactical, high speed data verification pipeline built explicitly to help suburban community coordinators manage the high stress environment of local neighborhood messaging chains during severe weather. When a conversational, unverified rumor hits a neighborhood group chat or social media feed, the block leader can paste the raw text statement directly into the dashboard. The application immediately processes the claim and matches it against current official infrastructure data to output a clear safety verdict. A green ‘Confirmed’ status indicates a definitive spatial match with an active National Weather Service alert. Conversely, a red ‘Contradicted’ status indicates that live system parameters explicitly disprove the claim, such as the ERCOT grid dashboard showing 100% stable operating reserves during a rolling blackout rumor. If the system detects inconclusive data or overlapping timelines, it marks the claim as a yellow ‘Unverified’ status to signal that the user should proceed with caution. Previously, managing group chats during a crisis meant trying to decrypt complex radar data when you may have no experience. Instead of forcing the block leader to interpret technical coordinates in the middle of a storm, VerifyDFW generates a pre-formatted message complete with official sources. With a single click, the coordinator can copy this analysis and drop it right back into the local community text thread, thereby stopping viral panic before it can spread across the neighborhood.

How we built it

For the engine, we used Meta’s Llama 3.1 8B model served through the Groq API because we needed low latency text processing. When a user drops a conversational message (or Twitter/X URL) into the app, the model handles the natural language understanding. We used strict Pydantic data contracts in Python to enforce NER (named entity recognition) on the LLM's output. This forces the model to reliably extract two critical data keys from raw human text: the implied location entity and the specific emergency claim type (like a tornado, flash flood, power grid hazard, etc.), returning them as a clean JSON schema. Once our FastAPI backend receives this metadata, it routes it to our data verification pipeline. For environmental threats, the backend fires off concurrent and asynchronous requests to live public safety streams, including the National Weather Service Alerts API and USGS Stream Gauges. Because public utilities and local municipal law enforcement do not expose open APIs for consumer queries, we instead used robust mock layers for our ERCOT grid capacity indices and local Police Department dispatch logs to simulate live state changes and validate our cross referencing engine. For the frontend dashboard, we built a minimalist user interface designed to be highly readable under pressure. To handle cases where a forwarded rumor doesn't contain a specific city or neighborhood name, we built a geolocation fallback mechanism. The frontend React application captures the coordinator's preset browser coordinates or home zone profile and appends that metadata directly to the backend request headers, allowing the verification engine to anchor its data tracking footprint automatically.

Challenges we ran into

Our primary software hurdle was that disaster rumors arrive as conversational text strings rather than database fields. We had to spend significant time iteratively making system prompts and designing strict Pydantic schemas to force our Llama 3.1 parser to reliably extract accurate location parameters and standardized threat types without completely hallucinating. Additionally, we faced severe data collection barriers because neither ERCOT nor municipal police departments expose open APIs for consumer applications. We had to pivot and develop realistic mock backends to simulate grid load fluctuations and emergency dispatch logs so we could properly test our cross-referencing logic. Lastly, handling text inputs that lacked any explicit location metadata, such as a forwarded text simply stating "the power just cut out,” proved difficult. We solved this by creating a fallback system that captures frontend browser coordinates and cross-references active county-wide NWS warning polygons to anchor the search footprint when the raw text is missing context.

Accomplishments that we're proud of

We are incredibly proud of designing a platform that fundamentally changes the safety of neighborhood messaging networks. We hoped we would be able to give local volunteers a functional shield against viral panic during critical times, and we think we accomplished that with this project. Furthermore, we are proud of building a safety override that completely eliminates the risk of a model hallucination causing a dangerous false negative. VerifyDFW is strictly barred from ever contradicting an active tornado or flash flood claim. Even if live official API feeds are completely silent, those high stakes verdicts are forced to stay at 'Unverified,' ensuring that our technical implementation never comes at the cost of human safety.

What we learned

Building this project taught us that the true impact of applied AI in public safety depends entirely on a human-in-the-loop architecture. We learned that AI should never try to act as an autonomous emergency dispatcher; we initially thought that we could have the AI also send out the messages, but we realized quickly that this would be a very poor decision. We also realized that VerifyDFW is most impactful when it respects existing local human networks rather than trying to bypass them. By focusing our tool on arming a single, trusted community gatekeeper with a fast, sourced analysis instead of automatically blasting out public alerts, we preserve vital human judgment and localized expertise. We learned how to build software that assists human leaders during a crisis, ensuring they can protect their neighborhoods without being forced to make safety decisions blind.

What's next for VerifyDFW

Moving forward, we want to expand VerifyDFW's localized ingestion by building automated scraper webhooks directly into regional municipal public safety CAD (computer aided dispatch) logs. This will allow the platform to instantly verify local and street-level utility incidents, water main breaks, and also down lines the moment local fire and police units are dispatched, thus keeping neighborhood block leaders steps ahead of the crisis. We also plan to integrate with Oncor’s utility system architecture to transition away from our engineered mock layouts and stream actual localized outage data. This will allow VerifyDFW to instantly cross-reference a neighborhood rumor about blown transformers or street-level grid failures directly against live line diagnostics and active crew ticket data inside the DFW footprint