Day 3: Forcing AI Neutrality—Architecting a Zero-Hallucination Meteorological Engine on Poe
Hey IH, Day 3 of the SkyHound Weather Sniffer launch.
For those who missed it, my vision for Loretta Compliance goes way beyond just pulling weather data. I founded this company to build automated, regulatory-compliant AI architectures that bridge the gap between contractors and insurance carriers. The core mission is to end the subjective tug-of-war that constantly uses the homeowner as the rope—tackling everything from multi-trade building code validation to B2B forensic scope auditing. Neutrality holds no opinions, and neither should our systems.
SkyHound—our meteorological tracking module—is just the first piece of that puzzle. Today, I want to dive into why I deployed this specific engine through Poe, our dual-engine architecture, and exactly how we force an LLM to accurately sniff out swath coverage probability without hallucinating.
The Economics of Deploying on Poe
Running deep-research meteorological queries via standard API wrappers gets expensive fast. By structuring SkyHound as a prompt-based bot on Poe, we shift the compute economics entirely.
Instead of burning our own API budget, users leverage Poe’s ecosystem, utilizing their own subscription point allowances (like the $19.99/mo plan for 18,000 points) to run the complex queries. Furthermore, Poe’s creator monetization allows us to set a per-message price in USD. The field reps get deep research capabilities, context retention, and automated reports, while we maintain a highly scalable, low-overhead distribution model and earn revenue per message.
System Architecture: The Dual-Engine Fallback
Since AI natively wants to chat and output conversational filler, I had to build a rigid behavioral engine with a built-in logic loop breaker. The bot is designed strictly as an Adaptive Telemetry Report Generator.
To ensure it works anywhere, we implemented a Dual-Engine Capability Detection system:
Mode A (Sandbox Native): If the environment supports Python execution, it silently runs native libraries (like hashlib) to generate a SHA-256 cryptographic audit stamp and renders our HTML schema directly into a downloadable PDF.
Mode B (The Poe/Claude Fallback): If the bot detects it is running on Poe, Claude, or Grok—or if code execution fails—it instantly aborts the code without displaying errors to the user. It seamlessly falls back to a clean Markdown schema, adding simple instructions at the bottom for the user to print/save as a PDF directly from their mobile device.
The "Bite-Guard" Firewall (UPPA Compliance)
Building for contractors requires navigating strict Unauthorized Practice of Public Adjusting (UPPA) laws. The system is legally firewalled. The LLM is explicitly barred from interpreting insurance policies, discussing building codes, or diagnosing physical property damage—it can only confirm that a weather event occurred.
Furthermore, we enforce strict PII scrubbing. While street-level addresses are ingested for probability mapping and generating the hash stamp, they are immediately purged and never output in the final visual report (displaying only City, State, ZIP).
The Edge: Swath Probability & Active Geospatial Retrieval
Standard AI wrappers hallucinate impact locations. We enforce a "Zero Hallucination Lock" that strictly forbids estimating, interpolating, or inventing weather data.
Instead of hallucinating, the prompt logic mandates an Active Geospatial Data Retrieval sweep across live NOAA, NWS SPC, and historical radar archives for the last 3-5 years. If no direct spotter pins exist for the immediate target, it triggers a grid-expansion fallback, searching the surrounding 5-10 mile county/NWS radar grid. It evaluates all recorded events against standard carrier thresholds (e.g., Hail >= 1.00", Wind >= 58 mph).
The AI calculates the Swath Probability—a statistical percentage from 0% to 100%—that the historical storm system covered the exact target coordinates based purely on the retrieved swath data. It then forces the LLM to output the exact phrasing: "[X]% actual chance of target coordinate storm coverage" alongside the cryptographic verification stamp and source attribution.
By anchoring the AI directly to verified sweeps and forcing a rigid output schema, we deliver undeniable facts. Good dogs follow the rules.
