03 — Deliver · Product group

IGNITE:AI

The intelligence layer

IGNITE:AI is the software layer between perception and action. It runs your trusted AI models at the edge — unchanged, in isolation — and delivers georeferenced findings straight into existing command systems. No retraining, no dataset sharing.

Request a briefing Download the IGNITE:AI documentation
TRACK 014TILED TO YOUR MODELTRACKS OUT, NOT VIDEO

See IGNITE’s features ↓

  • AI pluginDeploy your trusted models at scale — PyTorch, TensorFlow, ONNX or OpenCV DNN, converted to TensorRT and run unchanged in an isolated container.
  • Sized for the Jetson16 GB, shared between the pipeline and your model.
  • Precise georeferencingEvery finding placed on the ground using terrain and platform angles, held across frames.
  • EncryptedModel weights and mission data stay in the encrypted store on the platform. Zero Sapient access.
  • Open outputsFindings out as Cursor-on-Target (CoT) into ATAK, plus JSON detection reports and annotated JPEG crops — into existing C2 / C4, not video.

The intelligence layer

The Layer Between Perception and Action

ROOK or ECHO capture, FORGE computes, and IGNITE:AI runs your trusted model at the edge — unchanged, in isolation. Bring your own model, connect third-party sensors and external feeds, and run on our compute or yours.

IGNITE:AI platform architecture The IGNITE:AI framework at the core of the Sapient platform. AI models from three sources (your own, a partner's, or Sapient's) and sensor and data inputs (Sapient ROOK and ECHO cameras, third-party cameras, external feeds such as AIS) feed into IGNITE:AI, which runs on Sapient FORGE by default or the customer's own NVIDIA hardware as an option. SAPIENT PERCEPTION IGNITE:AI — PLATFORM ARCHITECTURE AI MODELS Bring your own, a partner's, or ours Your own model DEPLOYED BY YOU Partner model FROM A SOLUTION PARTNER Sapient model DEVELOPED BY SAPIENT SENSORS & DATA Connect & fuse multiple inputs Sapient cameras ROOK & ECHO · 10K OPTICS Third-party cameras MULTIPLE TYPES & MODALITIES External data feeds AIS & OTHER SOURCES IGNITE:AI The intelligence layer — your model runs unchanged, isolated & encrypted ONBOARD · AT THE EDGE · ZERO SAPIENT ACCESS RUNS ON DEFAULT Sapient FORGE EDGE COMPUTE · INCLUDED OPTIONAL Your NVIDIA hardware BRING YOUR OWN COMPUTE
  • Open at every edgeYour own model, a partner’s or ours; Sapient and third-party sensors; external feeds; FORGE or your own compute.
  • Secure at the coreModel weights and mission data stay in an encrypted store on the platform — onboard, at the edge.
  • Your IP, your controlSapient has zero visibility into your models, training data, or operational output.

How it works

One Frame in. Compliant Intelligence Out.

Three modules wrap your AI model: optimise each 10K frame for the model, run inference in an isolated container, and turn detections into usable, C2-ready products.

Input module

Optimise the Data for the Model

  • Prepare the raw 10K frame — normalisation and calibration before inference.
  • Scale and tile to the model’s input size, GSD-optimised, so every model sees correctly-scaled data.

Inference module

Your Trusted Model, in a Container

  • Runs unchanged in an isolated Docker container — framework-agnostic, no retraining, no dataset sharing.
  • Detection and classification on tiles (e.g. TensorRT) return classes and bounding boxes.
  • Update or swap models as a container swap, not a rebuild — run several in parallel.

Output module

Turn Detections Into Usable Products

  • Extract high-res crops — return to the raw 10K frame and crop each detection at full resolution.
  • Georeference and track — precise geolocation per finding from terrain and platform angles, IDs held across frames.
  • Conform to C2 — package as Cursor-on-Target (CoT) for ATAK and existing C2 / C4.

IGNITE:AI · Capabilities

Capture. Process. Deliver.

The software layer between perception and action. Core capabilities ship with every IGNITE:AI deployment; add-on modules extend it for specific interoperability and mission needs. Select a feature for detail.

Feature Core Add-on
01

Capture

Smart triggering — speed, altitude, position, time

Captures on mission-relevant conditions rather than a fixed clock — ground speed, altitude, position, and time interval, alone or combined — so frames are acquired where and when they matter.

Hybrid capture mode — H.265 video with detection-grade stills

Runs a H.265 video stream alongside full-resolution still capture at defined intervals. The model runs on the detection-grade stills and overlays its findings onto the live video — continuous video for the operator, detection-quality frames for the model.

Smart Sync — time-aligned multi-sensor capture 3rd-party · roadmap

Synchronises multiple Sapient sensors to a common time and frame reference for aligned multi-sensor capture. Synchronisation with third-party sensors is on the roadmap.

Multi-region-of-interest streaming — parallel ROIs from one 10K frame

Streams several independent regions of interest from a single 10K frame in parallel, so distinct areas can be worked simultaneously without separate passes.

02

Process

Model-aware scaling & tiling — GSD-optimised input

Prepares the 10K stream to fit your model — normalisation, calibration, and GSD-optimised tiling sized to the model's input — so inference runs on correctly scaled data.

Framework-agnostic inference — any framework, isolated container

Runs your model unchanged in an isolated container, independent of the AI framework it was built in. Updates are a container swap, not a rebuild.

Multi-model mode — parallel models, selectable in flight

Runs multiple models on one sensor stream in parallel — for example detection alongside vision-based navigation — with the active model selectable in flight. Swap or add models as threats and environments change.

Frame-to-frame tracking — persistent IDs across frames

Tracks unique findings across frames so a detection persists as a single object rather than re-reporting each frame.

Context-aware georeferencing — terrain and platform angles

Places findings on the ground using terrain, gimbal and platform angles, and related context, rather than a flat projection.

Geofencing — bound activity to an area of interest

Constrains capture, processing, and reporting to a defined geographic boundary — focus on an area of interest, or exclude sensitive and out-of-scope zones.

Sensor fusion — multiple inputs, one aligned product

Combines multiple sensor inputs into one aligned product, including alignment of stored data.

Model & data security — encrypted, isolated, zero Sapient access

Model weights and mission data are held in an encrypted store on the platform; inference runs in isolation. Both model and data are protected at rest, with zero Sapient visibility.

03

Deliver

Cursor-on-Target → ATAK — georeferenced findings into C2

Emits georeferenced findings as Cursor-on-Target messages into ATAK and existing C2 workflows — intelligence out, not video.

STANAG interoperability — conform findings to STANAG roadmap

Conforms findings to STANAG for interoperability with existing C2/C4. Delivered as a paid add-on module; on the roadmap.

Bring a Model. Leave With an Intelligence Pipeline.

We will walk your existing AI into the IGNITE:AI container format against a live 10K stream.

Request a briefing How to integrate