Pre-revenue · ready to pilotOpen demos — no login wall. Authenticated pilots at gprkinetic.pro/os.

Target ballast spendby chainage. Not corridor.

Kinetic GPR ingests GPR surveys through Theorem — our inference core — and returns chainage-locked fouling maps and Selig Fouling Index (FI) proxies. Track engineers sign off on clean, fouled, and marginal ballast calls; asset managers get a maintenance scope they can put in front of procurement.

Diagnostic twin

Open →Open →

Economic Oracle

Compare blanket vs targeted CapEx

Fouling and moisture cluster by chainage. Most corridors do not need uniform ballast replacement. Adjust length and unit cost below; defaults follow a GTA-style structure from public GPR literature.

Model assumptions

This is an illustrative scenario tool, not a quote or guaranteed CapEx outcome.
Defaults use a GTA-style reference structure (corridor length × unit cost × intervention %). Actual savings require operator ground truth (cores, lab Selig FI, program rates).
Intervention % is a modeling input (live Foundry rows when available; otherwise ~39% reference).
Pre-revenue Kinetic OS — pilot-ready. Do not treat dollar figures as measured customer results.

Corridor length (km)

Renewal cost ($M / km)

Blanket renewal

$18.0M

100% corridor replacement assumption

Targeted intervention

$7.0M

38.9% of chainage · GTA commuter reference

Scenario capital differential

$11.0M

GTA reference structure · your inputs recalculate live

Request pilot briefing

Capability transparency

Current state vs research vs vision

Kinetic OS is pilot-ready and pre-revenue. Strong rail language describes the product direction — the tiers below say what is reliable now versus still research.

TODAY

Works reliably today

Hardware-agnostic ingest paths (DZT / SEG-Y / SRS parser) with Real Data Only gating
Human-in-the-loop labeling → geometry_json annotations
Deterministic scoring hooks + Supabase-backed dashboards (pilot demos)
Illustrative Economic Oracle (adjustable parameters — not a customer result)

PILOT

Active research / pilot

Res-SAM / reservoir segmentation on annotated real corridors
Foundry analytical ontology (FoulingMeasurement, InterventionScope — pilot/research seed)
Multi-sensor fusion (RTK + encoder → chainage) and TensorRT edge path
Public CMU-GPR / open-rail demos labeled as public data

VISION

Future vision

Full autonomous subsurface digital twin with zero human review
Continuous Pulse fleet-scale telemetry on operator lines
Operator ground-truth CapEx outcomes (cores + production BFI calibration)

Inference core · pilot path

Theorem

Subsurface certainty. Delivered.

Theorem is the inference and ontology layer behind Kinetic GPR — the neocortex that turns radargrams into chainage-locked ballast calls your track engineers can sign off on.

Stop interpreting in isolation. Start knowing by chainage — with a semantic rail model built for pilot corridors today and production ontology workflows next.

Scan ingest
Vendor-native GPR (.DZT, SEG-Y) aligned to chainage
Semantic ontology
Track, corridor segment, and ballast objects — not flat CSV rows
Theorem inference
Res-SAM segmentation with engineer sign-off before commit
Proof surface
Diagnostic twin and CapEx scenarios asset managers can defend

Theorem · inference core

Res-SAM · ontology

Ingestion pipeline

Raw acquisition to verified anomaly

Upload vendor-native radargrams (.dzt, .rd3). The pipeline applies dewow, gain, and depth conversion, then runs Res-SAM segmentation with engineer sign-off before anomalies sync to the diagnostic twin.

Ingestion

Illustrative demo — the live pipeline runs in authenticated pilots.

Project Name

Scan ID

Hardware Origin

GPR Data Type

Drag and drop zone

Upload raw radargrams (.dzt, .rd3)

B-scan review

Scan: LBK-1004

0.0m2.5m5.0m7.5m10.0m

0m1m2m3m

AI model architecture

Production path: Res-SAM segmentation on B-scans plus structured regressors for Selig FI and permittivity. Legacy CRNN/XGBoost research metrics below are validation benchmarks on annotated corridor sets — not deployed edge models.

CRNN on B-scans

Hyperbola and layer-boundary patterns in radargrams — benchmark precision on held-out annotations.

Research val precision (not deployed)83.4%

XGBoost on A-scan features

Amplitude, phase, and attenuation vectors for fouling index and moisture estimation.

Val accuracy86.0%

Origin

Built by a track inspector. Scaled by ML engineers.

Kinetic GPR wasn't built in a lab. It was built by a former CN track inspector who saw how much ballast money was burned on blanket renewals without subsurface data.

Backed by Palantir's Founder Foundry

Pilot program

Request a corridor briefing

Share your corridor and role — we respond with a scoped demo and deployment outline. Open dedicated briefing page →