Per-observation
Each observation carries provenance, capture-time signatures, computer-vision tags, viewing-cone polygon, contributor tier, and a confidence score. Single observations never carry an incident on their own.
OpnIMG turns the first hours of a disaster into a structured stream of validated, geolocated damage observations. Tri-modal smartphone capture, a resilient five-path sync stack, multi-observer geometric clustering in the AWS cloud, and a partner-driven asset-resolution layer that hands clusters to utility, emergency-management, and 311 systems for downstream action.
Counterfactual modelling against ten historical storms in the day-3-to-7 window.
Four-factor (spatial · temporal · visual · semantic) similarity scoring.
LTE · Wi-Fi · SatCOLT · aid-station satellite · peer-to-peer mesh.
Eight provisional applications at the USPTO; four additional in confidential review.
OpnIMG emits validated, geolocated, computer-vision-tagged, viewing-cone-bound observation records — clustered when multiple observers converge on the same likely asset, and emitted to partner systems through documented endpoints. The utility OMS, the state EMA, or the 311 platform turns those observations into work orders and dispatches crews.
Each observation carries provenance, capture-time signatures, computer-vision tags, viewing-cone polygon, contributor tier, and a confidence score. Single observations never carry an incident on their own.
Clusters carry cone-intersection geometry (WKT), aggregated CV evidence, thumbnails, and a reference to an active event — storm, declared disaster, or self-derived implicit event.
Each cluster is emitted to the matching partner. The partner returns an opaque reference number — never an internal asset identifier — and OpnIMG records that reference in an append-only audit ledger.
Within the first 72 hours of a major event, the public switchboard collapses, satellite imagery is obscured by cloud or smoke, and field crews are sized for restoration rather than reconnaissance. Hurricane Helene took roughly seventy-one percent of Western North Carolina cell sites offline twenty-four hours after landfall. FEMA Public Assistance obligated more than four billion dollars in Category B emergency work for Hurricane Ian alone (DR-4673-FL).
The structural gap in that window is validation — too many candidate observations, too little assurance any of them are real. OpnIMG fills it by treating every smartphone and every responder device as a candidate observation node and applying geometry, computer vision, and cross-observer validation before anything is emitted downstream.
Counterfactual savings figures appear in OpnIMG materials only when labelled as model output, not as measurement. Operational figures from specific storms are not asserted in external materials unless a source-validated figure exists at the time of writing.
Each layer is engineered for the conditions of a disaster — degraded networks, intermittent power, unreliable GPS, adversarial submissions — and each fails safely toward the higher-trust state.
Three equally supported modes: an offline-first Progressive Web App sized for citizens and impromptu responders; native iOS with ARKit and LiDAR depth for crews; native Android on ARCore. All three write the same canonical envelope — image / clip, capture timestamp, GPS, compass heading, pitch, roll, device class, and a per-device capture signature.
Each submission passes through moderation, computer-vision labelling, geospatial reconstruction (cone polygon), H3 indexing, multi-observer geometric clustering, and composite confidence scoring. Cluster formation depends on cone-intersection geometry, CV class agreement, and temporal proximity — never on a single observation alone.
Clusters and validated observations exit through documented endpoints: signed webhooks, OGC WFS 2.0 for GIS tools, NIEM 6.0 JSON-LD for FEMA / DHS exchange, and an Esri Feature Service-compatible endpoint. The partner reconciles each cluster against its authoritative asset database and returns an opaque reference number.
The capture layer attempts, in priority order, the first available of five sync paths. Capture order is preserved across paths by signing every observation with its capture-time timestamp at the device — the backend always orders by capture time, not arrival time.
Standard LTE / 5G when the local network is healthy.
Open or municipal Wi-Fi where it is available.
Satellite-on-cells-on-light-trucks deployed by responding carriers.
Starlink and equivalent shared satellite uplinks staged at relief points.
Bluetooth or Wi-Fi Direct, store-and-forward to the next device that finds connectivity.
The platform consumes a layered set of authoritative storm and incident feeds so capacity is sized ahead of impact and every cluster is tagged with the right active event. Three tiers run continuously; a fourth is on the roadmap.
National Hurricane Center, polled every thirty minutes for active tropical systems — the authoritative source for declared hurricanes and tropical storms.
api.weather.gov, polled every five minutes for severe-weather, watch, and warning products — picks up tornadoes, ice storms, floods, and weather events outside the NHC envelope.
Implicit-event detector on a fifteen-minute rolling window — catches incidents that neither federal agency has yet declared, from the ground up. The always-on backstop when official feeds are quiet.
Radar fusion, deferred to a later release.
The active-events table is the system of record for which event a cluster belongs to. Partners can roll up validated observations by storm, by NWS warning, or by self-derived event — which is how Public Assistance documentation, mutual-aid sequencing, and after-action reviews get their event-level frame. When NHC issues an upgrade or NWS issues a warning, the gating layer pre-positions queue and Lambda capacity ahead of the surge; capture clients also receive a hint via /v1/config/disaster-status so the PWA can prioritise sync paths during the first window after impact.
Consolidation is not deduplication of similar pixels — it is geometric convergence on the same real-world asset, captured from different vantage points. The cluster engine groups observations when the geometry, the computer-vision interpretation, and the temporal window all agree. Observer identity is part of the weighting: higher-trust observers move clusters across the formation threshold faster than lower-trust ones.
Multiple observers viewing the same asset from different vantages — their viewing cones intersect in the real world.
The vision pipeline must read the observed scene the same way across observers.
Observations fall within the same active event window — storm, declared disaster, or self-derived implicit event.
Anonymous, phone-hashed, device-attested, and federated-SSO observers all contribute — higher-trust tiers carry more weight in opening a cluster.
A single observation — whatever its source — can never carry an incident on its own. A bad-faith submission cannot become a cluster by itself: an attacker would have to fabricate geometrically coherent observations from independent identities. Clusters can exist before any asset resolution; they carry their own geometric and visual evidence, and asset truth lives at the partner. The output to partners is a cluster — not a single observation — so the surface they consume is already validated.
An internal CEII, PII, and GDPR compliance review was completed in May 2026. That review — an internal opinion, not legal advice — produced the Reference-Number-Only Architecture that governs the platform today: it reduces our data-exposure surface to what the partner asset-resolution contract strictly requires, and explicitly removes the path that would have us holding an internal asset registry.
Cluster formation depends on cone-intersection, CV class agreement, and temporal proximity. No asset-match scoring against an internal asset catalog.
Clusters are bundled with polygon, aggregated CV evidence, and thumbnails, then emitted to the partner webhook. The partner returns an opaque reference number — never an internal asset identifier.
We do not centralise power lines, transformers, utility poles, substations, power plants, or primary distribution feeders. There is no internal mirror of partner asset data to protect.
The audit ledger stores the opaque reference number, not a partner_asset_id, asset_type, or asset_label. Per-observation retention drops from seven years to thirteen months; a separate aggregates table holds seven-year FEMA-style summaries.
Centralising utility asset data — particularly criticality flags and partner-internal asset identifiers — approximates the surface that FERC Critical Energy / Electric Infrastructure Information (CEII) under 18 CFR § 388.113 and NERC CIP-011 / CIP-014 are designed to restrict. Holding precise GPS, timestamps, and device metadata for many years also creates a larger PII surface under CCPA / CPRA and similar state laws than is strictly necessary for downstream partner action. This architecture removes the first exposure entirely and reduces the second to a defensible window. For partners with European data subjects, GDPR applies; OpnIMG is sequencing a documented lawful basis, a Data Protection Impact Assessment, and a Schrems II Transfer Impact Assessment with the 2021 EU Standard Contractual Clauses (with the UK addendum where relevant) before any European partner data crosses the platform boundary.
We are focused on the three buyer types where validated observations meaningfully change downstream decisions: storm and disaster response, utility operations (storm-mode and steady-state), and municipal 311.
State Emergency Management Agencies, FEMA regional staff, county OEM, mutual-aid coordinators.
Tier-1 investor-owned utilities in the U.S. Gulf region, public-power, co-ops. Storm-mode and steady-state.
Tier-1 city public-works, code enforcement, transportation, ROW management.
OpnIMG has not yet been deployed under a paid partner pilot. The company is identifying potential pilot candidates and is open to warm introductions to Tier-1 investor-owned utilities in the U.S. Gulf region, to state Emergency Management Agencies, and to municipal 311 system operators.
Operational figures from specific storms are not asserted in external materials unless a source-validated figure exists at the time of writing. Counterfactual savings figures — for example, modelled avoided losses across historical storms — appear in OpnIMG materials only when labelled as model output, not as measurement.
Eight provisional patent applications on file with the USPTO. Four additional applications in confidential review. The portfolio covers the methods that distinguish the platform — multi-observer geometric validation, viewing-cone polygon construction, cross-observer cluster formation, partner-driven asset resolution with audit-ledger semantics, and identity-tier-gated cluster trust weighting.
External materials always pair the two figures (eight filed and four in confidential review) and never describe the platform as relying on a single load-bearing patent.
Utilities, EM agencies, municipal CIOs, integration partners — use the form below. We do not publish email addresses.