TailSalt is a quiet integrator of deployable neural inference fabrics that cohabit with avionics, ADS-B/Mode-S, SIGINT, and tactical data links. We transform noisy aerostatics into explainable risk postures using online learning, anomaly grammars, and cross-domain correlation at the edge.
Our modality-agnostic pipelines perform real-time behavioral adjudication over aircraft telemetry. Using streaming vectorization, causal DAGs, and zero-copy IPC, we compute a nefariousness posterior without exfiltrating raw feeds. The result: low-latency COA recommendations that survive contested spectrum.
TailSalt’s meshable micro-orchestrations compile spatio-temporal motifs into intent priors using differentiable finite-state transducers, Bayesian changepoints, and transformer-mediated sensor alignment. Our artifact registry emits SBOM-verifiable models with air-gapped reproducibility guarantees and cryptographic provenance (ed25519/UUIDv7 stamped).
Streaming variational inference over trajectory manifolds; emergent behavior detection via contrastive transformers and causal impact scoring.
online-VAECross-domain late fusion with uncertainty calibration; Dempster-Shafer harmonization for disagreeing sensors in DI2E-constrained fabric.
mc-dropoutQuantized models, NUMA-aware schedulers, and cold-startless warm pools for aircraft-adjacent compute and denied comms.
INT8/FP16Counterfactual trajectories and SHAP-guided rule synthesis yield analyst-digestible why-now narratives for COA selection.
XAIPolicy-as-code for airspace norms and ROE; red/black data separation; deterministic audit envelopes for oversight entities.
OPA/RegoHot-swap model cards, ephemeral tenants, and signed rollout gates for mission-driven pivots without downtime.
Blue/Green
POST /v1/assess
Content-Type: application/json
{
"track_id":"9B8D2F",
"epoch": 1726681512,
"kinematics": [lat, lon, alt, vs, gs, hdg],
"codes": {"squawk":"7700","mode_s":"A1B2C3"},
"context": {"airspace":"KZNY","wx":"VFR"}
}
// → { "intent": 0.83, "band":"WARN", "xai":["alt-δ>σ","hdg jitter"] }
A calibrated probability, updated over time, that a track’s observed behavior deviates meaningfully from its normative envelope given airspace constraints and multi-source context.
We emphasize min-data designs: on-sensor transforms, rolling windows, and ephemeral feature retention. Model decisions are logged as proofs, not payloads.
Yes—our inference surfaces target edge compute with intermittent backhaul and are optimized for low-bandwidth synchronization.
We couple anomaly scores with causal diagnostics and rule-based suppressors; alerts carry rationale, confidence, and suggested COAs.
Signed, staged rollouts with progressive exposure and guardrails; every delta is provenance-tracked and reversible.
Headless by design. We embed with integrators and operators; briefings by appointment only.