On June 30, 2013, the Yarnell Hill Fire overran and killed 19 members of the Granite Mountain Hotshots in Yarnell, Arizona. It was the deadliest wildfire disaster for American firefighters since 1933. That same day, Bill King — the founder of Trajanus USA — was on mop-up duty at the Tres Lagunas Fire in Pecos, New Mexico, serving as a volunteer firefighter with Pecos Search & Rescue. Bill, his wife Jacue, and their son had been running night operations on that fire for weeks, living just one mile from the fire perimeter.

At the time, Bill was simultaneously working as a project manager in Santa Fe, designing GPS and RFID tracking systems for railroad operations. The collision of those two worlds — watching 19 firefighters die because nobody knew exactly where they were relative to a fast-moving fire, while spending his days tracking boxcars with sub-meter precision — produced the founding question behind Prometheus: "Why can't we track firefighters the way we track freight trains?"

Bill built the first Prometheus prototype in 2013, immediately after Yarnell Hill. In 2017, he presented it at the Esri User Conference with a live demonstration: a 350-mile field test generating over 15,000 GPS position reports with zero data errors. The prototype proved the core concept — real-time satellite tracking feeding into a GIS-based operational display — was technically viable and field-functional. Life shelved the project for several years. Now, in 2026, the AI and open-source landscape has caught up to the original vision in ways that weren't possible a decade ago.

Prometheus is a real-time wildfire intelligence and firefighter safety platform — a desktop application that unifies satellite GPS tracking, physics-based fire behavior modeling, machine learning prediction, and automated safety monitoring into a single operational tool for incident commanders. The platform is built on Tauri v2 with a Rust backend and React 18 frontend, using MapLibre GL JS and deck.gl for 3D terrain visualization. It is offline-first by design — every core feature works without internet connectivity, because fires burn in places without cell towers.

Satellite GPS tracking uses Garmin inReach Mini 2 devices communicating via the Iridium satellite constellation. Position reports transmit every 10 seconds from anywhere on Earth — no cell coverage required. The Portal Connect REST API feeds crew positions directly to the operational map. Meshtastic LoRa mesh networking provides backup communication at $30-37 per node with zero subscription fees, creating redundant positioning even when satellite connectivity degrades.

Fire behavior prediction is anchored to the Rothermel 1972 surface fire spread model — the same physics foundation used by every U.S. fire agency. Prometheus implements all 53 standard fuel models (Scott & Burgan 2005) through the emxsys/behave open-source JavaScript library in Phase 1, with a planned port to Rust in Phase 4. That Rust implementation will be the first of its kind — no Rust Rothermel exists anywhere in the open-source ecosystem. Fire spread propagation uses the Minimum Travel Time (MTT) algorithm with Dijkstra's shortest-path calculation, which is 100 to 1,000 times faster than the Huygens wavelet method used by legacy systems, and its grid-based approach maps directly to LANDFIRE's 30-meter fuel data resolution.

Real-time fire weather integrates through the Synoptic Data API, connecting to approximately 2,200 Remote Automated Weather Stations (RAWS) positioned at fire-critical locations nationwide. These stations deliver real-time temperature, relative humidity, wind speed and direction, and fuel moisture — the exact inputs the Rothermel model needs. The NOAA/NWS free API provides Red Flag Warnings and spot forecasts with GeoJSON polygon overlays for the tactical map. HRRR wind forecasts at 3-kilometer resolution drive animated wind particle visualization, showing incident commanders exactly where wind is pushing fire in real time. The Canadian Fire Weather Index (FWI) computes standardized fire danger ratings from weather observations.

Machine learning enhancement is designed as a complement to physics, not a replacement. An Attention U-Net architecture handles spatial fire perimeter prediction, trained on the National Dataset for Wildfire Spread (NDWS). A Transformer model captures temporal patterns — overnight recovery periods, wind shift timing, multi-day fire behavior evolution — that pure physics models systematically miss. The target performance metric is F1 ≥ 0.65 for next-day fire progression prediction. Estimated annual compute cost for GPU training and inference runs between $6,600 and $15,600, depending on model complexity and training frequency.

Terrain and fuel data come from two federal sources that are free and comprehensive. USGS 3DEP provides digital elevation models with 98.3% national coverage at resolutions ranging from 1 meter to 10 meters, all hosted on AWS S3. Slope, aspect, and elevation are critical fire behavior inputs — fire spreads faster uphill, and terrain channels wind. LANDFIRE delivers nationwide fuel model classification at 30-meter resolution, including fuel bed depth, loading, moisture of extinction, canopy cover, canopy height, and canopy bulk density for crown fire modeling. Together, these datasets give Prometheus the same terrain and fuel intelligence that government systems use, at zero data cost.

Automated safety monitoring is the heart of why Prometheus exists. The GeoLCES system automates the LCES safety protocol — Lookouts, Communications, Escape routes, Safety zones — the standard framework used on every wildland fire incident in the United States. The system continuously calculates escape route viability based on real-time fire proximity and spread prediction, triggers alerts when routes become compromised, and monitors safety zone distances for all tracked crews. In Phase 5, multi-crew tracking supports 20 or more simultaneous crews with ATAK/TAK interoperability for military-grade tactical situational awareness, IRWIN/IROC incident feeds, and full IAP (Incident Action Plan) data integration.

3D visualization brings everything together on a single screen. MapLibre GL JS renders terrain in three dimensions with deck.gl handling data visualization layers — fire perimeters, crew positions, weather overlays, wind particle fields. PMTiles provides offline tile storage so the entire map works without internet. USGS imagery and topographic basemaps are free and unlimited. Pre-computed arrival time surfaces from the MTT fire spread engine feed into WebGL shaders for 60 FPS fire progression animation on standard hardware.

The current leader in wildfire prediction technology is Technosylva, whose FIRESPONSE platform powers predictive services for CAL FIRE and other major agencies. Technosylva is effective but expensive, proprietary, and limited to agencies with the budget and infrastructure to support it. Federal programs like NOAA's Next-Generation Fire System (NGFS) and NASA's Digital Twin of Earth are tackling fire detection and atmospheric modeling at continental scale — important work, but focused on detection rather than the tactical tracking-and-safety problem Prometheus addresses.

Prometheus is not trying to replace these systems. It occupies a different position in the ecosystem: an affordable, offline-capable, open-source platform that combines tracking and prediction for use at the individual incident level. Where Technosylva serves state-level operations centers, Prometheus serves the incident commander in the field. Where NOAA and NASA solve detection, Prometheus solves the gap between detection and firefighter safety. The open-source fire behavior engine means the model is transparent and auditable — fire agencies can see exactly how predictions are calculated, unlike proprietary black-box systems.

Prometheus follows a five-phase build plan designed to deliver working capability at every stage. Phase 1: The Map (Weeks 1-4) establishes the desktop application shell with 3D terrain rendering, offline tile storage, and a mock GPS feed — a GPS dot moving on real terrain. Phase 2: Fire on the Map (Weeks 5-10) integrates the Rothermel engine with LANDFIRE fuel data and validates against the 2018 Camp Fire using the NIST TN 2135 dataset (2,200+ documented observations). Phase 3: Live Data (Weeks 11-18) connects real Garmin inReach GPS, live RAWS weather, NWS alerts, and HRRR wind forecasts into a functioning operational display. Phase 4: ML Enhancement (Weeks 19-28) trains the Attention U-Net and Transformer models and ports the Rothermel engine to Rust. Phase 5: ICS Integration (Weeks 29-40+) delivers the full GeoLCES safety system, multi-crew tracking, and ATAK/TAK interoperability for field-ready operations.

The total Year 1 build budget is estimated at $11,000 to $38,000 across all five phases, not including labor or drone LiDAR. This cost structure is possible because the stack is built almost entirely on open-source software and free federal data. GIS rendering is open-source (MapLibre, deck.gl). The fire behavior engine is open-source (emxsys/behave). Terrain data is free (USGS 3DEP). Fuel data is free (LANDFIRE). Weather alerts are free (NOAA/NWS). The primary recurring costs are Synoptic weather data ($500-2K/month), GPU compute for ML training ($6.6-15.6K/year), and Garmin satellite subscriptions ($30/month per device).

Prometheus is in active development with 10 completed research reports covering every component of the architecture, a locked technical stack, and a validated build plan. We're a small, experienced team — not a large organization. Bill brings 30+ years of federal project management experience with over $1.2 billion in managed programs, a PMP and CQM-C certification, and hands-on wildland fire experience. Tom Chlebanowski (PE, JD) provides licensed engineering oversight. The team has proven it can deliver validated technical systems — the TIA QA/QC Harness for Traffic Studies Engineering, with over 500 formulas and zero hardcoded outputs, was validated by Tom and locked on February 14, 2026.

We're actively seeking university research partners for the NSF FIRE grant program (April 1-7, 2026 deadline), fire agencies interested in future beta evaluation on real incidents, seed funding to accelerate Phase 3 and Phase 4 development, and wildfire science researchers who want to collaborate on open-source fire behavior modeling. We're building this with the community, not in isolation. Start small, scale together.