Sipsa Labs builds SIP: an on-device intelligence that lets a machine perceive its environment, decide, act, and learn from its own fielded data. No cloud in the loop. Our first application, Sentio, hears the drones that radio-frequency sensors structurally cannot.
Fielded hardware becomes a commodity. The durable asset is the intelligence on the device and the corpus of real-world data it accumulates.
SIP is a single perceive-decide-act-learn core that treats sensors as interchangeable frontends. The same runtime that scores an acoustic spectrogram can consume a vibration trace or a camera frame; the perception layer changes, the mind does not.
Every deployed node captures what it experiences, and that data flows into a versioned corpus that retrains the model behind a fail-closed gate: an update ships only when it provably beats the fielded model on held-out real data, with signed model packs and verify-on-load built into the release path. Machines that get smarter from their own service life, with an audit trail.
Autonomous waypoint drones and fiber-guided FPV platforms emit no radio signal. The deployed counter-drone detection base listens for radio. That gap is structural, and it is growing.
Sentio is an always-on fusion node. It hears the propulsion signature a drone cannot suppress, confirms it on camera, and reads the cooperative RF aircraft broadcast, Remote ID and ADS-B, so an aircraft that announces itself is identified and crewed traffic is explained off the alert queue instead of raising one. It decides on the device and publishes a bearing and track into the security picture an operator already runs (TAK-compatible today). It needs no spectrum license, it transmits nothing to detect the drone, and it runs on power-lean embedded hardware you can tile for coverage.
It is built as a detection and cueing layer that sits alongside the control-link RF and radar systems already deployed, covering the threat class those physically cannot hear. Thermal is designed into the sensor stage and is not yet fitted.
The same node is an instrument: every deployment captures labeled acoustic data from its own site, feeding the corpus that makes the next model better. The sensor is the entry point; the compounding corpus is the company.
| End-to-end prototype WORKING |
Acoustic cue, camera confirm, fused track, TAK output, live on embedded hardware. Bench-verified. |
| Detector performance HOLDOUT · LAB |
Pd ≈ 0.94 at ~1 to 2% false-alarm rate on a grouped held-out evaluation. A lab number, not a fielded one, and we label it that way on purpose. |
| Field characterization IN PROGRESS |
Controlled outdoor range campaigns with pre-registered protocols and control blocks. Results ship with methodology, under NDA. |
| Federal R&D IN REVIEW |
Multiple United States and allied research proposals submitted and pending. No awards yet; we say so. |
[HOLDOUT] or [FIELD]. If a number has no tag, we did not
publish it. Full evaluation methodology is available to qualified parties under NDA:
founder@sipsalabs.comDetection has a credibility problem: a decade of spec sheets that did not survive contact with the field. We engineer against that failure mode directly.
Every published figure states what it was measured on: held-out lab data, proxy simulation, or a real field campaign. The tags are not marketing, they are the release gate.
Thresholds, metrics, and exclusion rules are written down before the drone takes off. Control blocks are mandatory. If the protocol is violated, the gate refuses to certify the number, and it has refused ours.
When an approach fails our own evaluation, we retire it and document why. Buyers get the detection curve with its false-alarm analysis and failure accounting, not a highlight reel.
We would rather hand you a modest number that reproduces than an impressive one that does not. In a market of over-claimed spec sheets, being the vendor whose numbers survive independent test is the durable advantage.
Counter-drone sensing is the first application we are taking to the field, not the company. The same core extends to every machine that senses, decides, and acts.
The unchanged SIP runtime consumes vibration spectrograms and flags failing industrial equipment, validated on public benchmark data with zero counter-drone code edits. One mind, a second industry.
A per-plant sensing and precision-actuation platform for small farms, built on the same perceive-decide-act-learn core and the same honesty discipline: designed and bench-specified, with first measured perception baselines on public agricultural data.
Sipsa Labs is an independent, founder-owned engineering company in Colorado, building the full stack in-house: the detection models, the edge runtime, the data engine, and the physical nodes themselves, designed, printed, assembled, and bench-verified in our own lab.
We are early and we say so: a working prototype, active field campaigns, federal research proposals in review, and a fielded capture-and-retrain loop built end to end. If you run critical infrastructure or a facility with a drone problem, or you fund the machine intelligence layer this decade runs on, we would like to talk.
founder@sipsalabs.com