Library

SolvFilter

Hold the track picture when UKF stops scaling.

At fleet-scale state dimension, UKF prediction becomes the bottleneck — or it simply does not finish. SolvFilter keeps multi-target EKF prediction inside embedded budgets so counter-UAS and fusion programs keep a consistent track state instead of dropping contacts when the covariance step blows the frame budget.

Complete high-dimension predict steps where UKF times out or dominates wall-clock.
Validated deploy-survival narrative for radar tracking stacks — not a lab-only prototype.
Honest positioning: UKF alternative at scale; scipy EKF parity in calm noise, not a universal drop-in.
Fits Assured Sensing and counter-swarm architectures as the prediction layer after detection.

Product documentation

Multi-target track covariance propagation surviving at high state dimension on embedded hardware.

Imagine yourself in these moments. Same product, different industries.

Counter-UAV program

The track picture that lies.

A radar update arrives every tenth of a second. The fusion node must propagate dozens of coupled target states and full covariance matrices before the next measurement - on hardware that also runs detection, classification, and weapon assignment. When the prediction integrator blows up or drifts NEES-inconsistent, the kill chain breaks upstream. The operator sees a ghost track or loses a real one.

UKF is accurate enough at moderate dimension. At dim=140 it becomes the cost and consistency bottleneck - often timing out entirely. SolvFilter's validated pitch is as a UKF alternative at scale, not a drop-in replacement for every scipy EKF deployment.

Maritime tracking

The fleet picture at 2 Hz.

A surface-search radar refreshes on a two-second cadence. Forty contacts, each with a 12-state kinematic model and a full covariance — the predict step has to finish before the next dwell. UKF cost scales until the fusion node simply skips cycles.

SolvFilter is built for that predict step: Riccati propagation that stays inside the frame budget when UKF does not.

Test range replay

The regression that only fails on hardware.

Lab replay passes. Field tape fails. The difference is dimension and update rate, not the physics model. A tracking stack that validated at dim=24 silently degrades at dim=140.

Validation packs (RCT/MTTA) exist so you can replay the same scenario class before you promise deploy survival on a program slide.