Presented on FRINGE 2026
Operated by AI InSAR-Specialist Agents, with the Human Expert in Command.
AI InSAR specialists are on call at every gate of the processing chain. Specialized agents read live metrics, configure parameters, recover from failures, and operate the workflow autonomously while escalating ambiguity and potential deformation to the expert.
Book a demo
Schedule a 30-minute demo
If the inline scheduler does not load, open Calendly in a new tab .
Elastic orchestration for stage 5. The agents command an execution band that schedules work, fans out short-lived containerized workers, and writes outputs to shared storage.
Elastic execution band
scales with the pool - no design limit
Autonomous end-to-end by default: specialized agents own the gates; the expert enters only at the handover triggers.
within guardrails, high confidence
applies + logs rationalescience-altering choices
evidence + one-click expert validationambiguity or possible deformation
evidence dossier to the expertGUARDRAILS : Code-enforced: network must stay connected / flag, never delete / possible deformation always goes to the human / every decision logged.
Generates candidate interferogram networks (six graph strategies x tunables - candidates cost nothing), scores connectivity, redundancy and decorrelation risk; commits the best and re-shapes it after first coherence results.
Hands over: tied candidates / ambiguous monitoring objective.
Reads per-pair coherence fractions and the network graph; drops dates that degrade the stack but keeps "noisy bridges" that hold the network together.
Hands over: seasonal coherence loss / network fragmentation.
Reads phase dispersion, closure residuals and connectivity; places the discard cutoff on the error tail, and re-runs failures in the cloud under a new configuration.
Hands over: ambiguous cutoffs / date-wide failures.
Reads phase-elevation correlations, patch coverage, stable-point ensembles and before/after residuals; re-runs correction stages with scene-tuned parameters.
Hands over: whenever residual signal could be deformation.
The agents command an elastic cloud back-end for the heaviest stage - phase unwrapping. Each interferogram unwraps in its own short-lived containerized worker: no shared state, embarrassingly parallel, disposable by design.
the stack fans out: one worker per interferogram, as many workers as the pool provides. Capacity grows with the pool, not the code.
each worker is sized to its task: machine class and unwrapper tiling chosen per interferogram by the scheduling agent.
turnaround collapses from multi-day sequential runs to a fraction on an elastic pool.
retries, per-task timeouts, multi-zone failover, self-deleting workers, crash recovery: a run survives its failures.
There is no architectural ceiling in this design. The workers are stateless and disposable, and the same orchestrator can command ten workers or ten thousand. Scaling is therefore a provisioning decision, not a design change.
Representative artifacts generated across the SBAS chain, illustrating how automated quality control, expert oversight, and continuous validation are applied throughout the processing workflow.
The SBAS chain is built on algorithms established in the published, peer-reviewed InSAR literature. Our contribution is the framework around those methods: a deep-context knowledge corpus, specialized AI agents that operate the chain autonomously while keeping the expert in command at every gate, and an elastic, vendor-neutral back-end that scales without requiring design changes.
The SBAS chain is built on algorithms established in the published, peer-reviewed InSAR literature. Our contribution is the framework around those methods: a deep-context knowledge corpus, specialized AI agents that operate the chain autonomously while keeping the expert in command at every gate, and an elastic, vendor-neutral back-end that scales without requiring design changes.
Every piece of knowledge the chain depends on is written down where the code lives: sensor data formats and their pitfalls, algorithm specifications distilled from the published literature, the threshold semantics of every quality gate, and expert feedback together with the reasoning that resolved it. The result is a structured, version-controlled context corpus that grows with the codebase.
The same corpus serves twice. At build time it made AI coding assistants domain-aware, and a 3-person team comprising an InSAR specialist, a geologist, and a software developer brought the chain to operational readiness in about seven months. At run time it becomes the knowledge base of the specialized agents that operate the pipeline: each agent loads the context of its own stage, reads the live metrics, and acts.