Every place has a second geography — the one written in radio. We are building the models that read it, and the platform that watches it continuously: fusing live spectrum, signal propagation, and physical terrain into one answer to what is transmitting, from where, and whether it should be.
Radio became load-bearing infrastructure while the tools for observing it stayed manual, episodic, and expensive. We think that gap closes with models that learn from measured spectrum while staying bound to the physics and the ground beneath it.
Timing for power grids, financial networks and cellular backhaul arrives from satellites transmitting at a power level that reaches the ground weaker than the thermal noise around it. Ports, mines, farms, hospitals and factories now run private LTE and 5G. Industrial telemetry, building automation, medical telemetry and uncrewed aircraft all share unlicensed spectrum with every consumer device in range. The radio environment has quietly become a dependency of physical operations — and, unlike the network stack above it, it is almost never instrumented.
When something goes wrong there, the response is still a specialist with a spectrum analyzer, a truck, and a day. That is a drive-test answer to a continuous problem: it tells you what the environment looked like on Tuesday afternoon at one location, which is exactly the wrong shape of information for interference that is intermittent, mobile, or deliberate.
Two mature toolkits exist and neither closes the gap. Deterministic propagation modeling — Longley-Rice, ITU-R recommendations, ray tracing — encodes real physics, but describes a static, cooperative world you have to specify by hand, and it does not learn from what actually happened. Modern spectrum machine learning learns readily from data, but is typically terrain-blind: trained on captures divorced from the geography that shaped them, so a classifier that works beautifully at the site it was trained on degrades at the next valley over, and nobody can say why.
The useful system is the fusion of the two. A model should be free to learn what an emitter looks like, while being told what the ridge line, the tree cover, the building mass and the antenna geometry make physically possible. That constraint is not a limitation — it is most of the signal.
Concretely: inexpensive distributed sensors stream spectrum to a service that continuously answers three questions — what is transmitting, where is it, and is that normal here? The platform is how the answer gets delivered. The terrain-conditioned models underneath are why the answer is worth anything.
The three disciplines below are not three product lines. They are one problem, and treating them separately is precisely what makes existing tools brittle.
Elevation, clutter, landcover and built environment folded into propagation-aware models, so a prediction understands the ridge line, the canopy, and the tower it has to get past. Every observation is joined to its physical context before a model ever sees it — the terrain is a feature of the data, not an afterthought in the analysis.
Learned models over raw spectrum: detect an emitter, characterize what it is, place it geographically, and track how its behaviour changes over time. Geolocation combines multi-node time-difference and received-power estimates with a propagation prior, rather than assuming the free-space model that the real world never obeys.
The wireless attack surface is a physical one. We study how these systems fail — spoofing, jamming, interference, denial — and build the monitoring that notices first: a baseline of what normal looks like at a specific place, and an evidence trail an operator can act on and defend afterwards.
Site owners whose operations depend on radio they do not currently observe: critical infrastructure and utilities, ports and logistics hubs, airports and heliports, industrial campuses running private cellular, and the security teams responsible for all of the above. Today they discover interference from its consequences. The premise of this company is that they should discover it from a monitor.
Spectrum monitoring is a data-volume problem before it is a machine-learning problem. The architecture below is what the physics requires: aggressive reduction at the edge, cheap durable storage for the raw captures worth keeping, a warehouse where every observation can be joined against terrain, GPU training on the resulting corpus, and inference pushed back out to the sensors.
Low-cost software-defined radio nodes with a small compute board. Channelization, energy detection and first-pass classification run on the node, so what leaves the site is detections, features and selectively triggered raw capture — never an unfiltered firehose.
Detection events stream continuously into a message bus; triggered raw IQ lands in object storage as immutable, checksummed captures. The split matters: the event stream must stay real-time, while the raw archive exists to retrain models against evidence months later.
Each detection is enriched with the physical context of where it was heard: elevation profile, landcover, building mass, and a computed propagation prior over candidate emitter locations. This is a geospatial analytics workload as much as a signals one, and it is the step that makes the models transferable.
Representation learning over unlabeled spectrograms and IQ, then task heads for emitter classification, localization and anomaly scoring — each conditioned on the terrain features from the previous stage. This is the GPU-bound part of the system and the one that scales directly with available compute.
Compact models are pushed down to the sensors for sub-second local detection, while a cloud tier fuses across nodes for geolocation, keeps the historical baseline of what is normal at a given site, and drives alerting and the operator console.
Every candidate model is scored against replayed interference, jamming and spoofing scenarios, and against held-out geographies it has never seen. A model that has not been attacked has not been evaluated — in this domain that is a security property, not a benchmark.
Those numbers are the whole constraint. A serious training corpus means petabyte-class object storage and a warehouse that can join signal features against terrain rasters without a bespoke cluster; a serious model means sustained GPU training rather than opportunistic single-card runs; a serious pilot means inference serving that scales to zero between events. For a small pre-revenue team, cloud capacity converts almost one-to-one into experiments actually run — which is why we are building on managed infrastructure from the start rather than deferring it, and why we are talking to cloud and compute partners this early.
Learned systems that ignore the physics they sit on top of are confident in exactly the situations they should not be. Ours are built the other way around.
Real captures, real terrain, real hardware. Synthetic data is for augmentation and for stressing edge cases, never for the claim that something works.
Propagation, geometry and sensor limits are constraints on the model, not features it is left to rediscover — the reason a result at one site should be expected to hold at another.
Everything is evaluated against someone actively trying to defeat it, because in this domain someone eventually is — and against geography the model has never seen.
Smoldering Palms is a small engineering team, pre-product and pre-revenue on this work. Everything below the current phase is stated as intent, not as something already shipped. The sequencing is the plan: build the ground truth before the model, the model before the product, the product before the pitch.
Capture rig, collection methodology, and reproductions of the classical propagation baselines against our own measurements — so every learned model has an honest, physics-derived thing to beat. Deliberately no customers and no pilots at this stage; a monitoring claim that has not been measured is a liability.
A multi-node sensor testbed over one real geographic area, running continuously into the ingest pipeline. The output is the asset that does not currently exist publicly at this resolution: a labeled, terrain-joined corpus of live spectrum with known ground truth about the emitters in it.
Terrain-conditioned models for classification, localization and anomaly detection, trained on that corpus — with multi-node geolocation where the propagation prior does the heavy lifting instead of a free-space assumption. This is the phase the GPU budget belongs to.
Pilot deployments at sites that already have a spectrum problem they cannot see: a monitoring console, alerting, and assurance reporting a site owner can actually run a shift from. Product decisions get made here, against real operators, rather than in advance of them.
Deploy into terrain the models have never seen, with little or no local labeling. This is the real test of whether conditioning on physical context bought anything — and, if it did, the point at which the system stops being a deployment and starts being a product.
The RF work is early; the engineering practice behind it is not. The operations platform behind this site's login is production software we designed, shipped and run ourselves — a full-stack application on edge infrastructure with its own database, migrations, automated test suite, nightly backups and identity-gated access, deployed continuously. We build systems that stay up and stay auditable, and we intend to hold the spectrum work to the same standard.
If any of the above is your problem, your research area, or your infrastructure, we would like to hear from you. Concretely, four conversations we are actively looking for: