Multimodal maritime intelligence foundation model

The maritime intelligence model.

Vanchi turns fragmented maritime signals into evidence, assessments and action — across detection, investigation, collection and reporting.

Request Access See Vanchi Reason
Case VN-0417 · Dark vessel investigation Simulated demonstration · times Zulu
04:10ZTrack established. Cooperative AIS, southbound in corridor.
09:42ZAIS gap opens. No fix for 8 h 14 m in a reporting-dense area.
17:56ZSAR detection. 31 nm beyond dead-reckoned position.
18:03ZRF observation. Navigation radar radiating on a consistent bearing.
18:04ZAssociation proposed. Detections linked to the gapped track.
18:04ZAssessment drafted. Evidence, contradiction and alternatives attached.
18:21ZAnalyst review. Association accepted. Collection requested.
18:22ZCase updated. Timeline, tasking and draft report revised.

Assessment — likely continuation of gapped track

Confidence0.74 · simulated
  • SAR extent consistent with vessel class
  • RF bearing agrees with projected course
  • No VIIRS return on overhead pass

Recommended collection: SAR revisit of the projection area, next window 21:14Z.

Simulated case with mock data. Confidence values are illustrative, not measured performance. DETECT → RANK → EXPLAIN → REVIEW → TASK → CONFIRM → REPORT → LEARN

The problem

The ocean does not produce a single source of truth.

What a vessel is, and what it is doing, is scattered across tracks, imagery, registries, sensors, ownership records and analyst judgement. These sources disagree — routinely, and sometimes deliberately.

Most systems display each source in its own pane and leave the reconciliation to the analyst. Vanchi is built to reason across all of them as evidence: weighed, cited and held open where they conflict.

Product proof

See Vanchi investigate.

Three example cases, worked end to end. Hover any underlined claim to see exactly which evidence supports it — and which contradicts it. Simulated demonstrations

Triggering observation · AIS gap, 8 h 14 m

A vessel stops transmitting inside a reporting-dense corridor.

Vanchi searched the projection area and found a SAR detection consistent with the vessel's length class, supported by an RF observation on a bearing that agrees with the projected course. The gap itself is atypical for this route, where reception is near-continuous. One source disagrees: the overhead VIIRS pass returned nothing at the projected position, which is consistent with low-light conditions but is retained as contradicting evidence.

Assessment confidence 0.74 Supporting 3 Contradicting 1

Alternative hypotheses — retained, not discarded

  • Transponder fault, same vessel p 0.15
  • Different, AIS-exempt vessel p 0.08
  • Reception gap only p 0.03
Accept associationRejectRequest collectionAdd note

Evidence

E-201SAR detection, 17:56Z — extent 178 ± 14 m, 31 nm from last fix
E-202RF intercept, 18:03Z — X-band navigation radar, bearing 041°
E-203Historical reception baseline — corridor coverage 98.7%
E-204VIIRS pass, 17:31Z — no detection at projected position

Recommended next step

Task a SAR revisit of the projection area at the next window, 21:14Z. Expected to raise or collapse the leading hypothesis.

Draft report excerpt · evidence-linked
A vessel assessed as the continuation of gapped track T-8841 [E-201, E-202] is proceeding northeast outside its declared route. One contradicting observation is recorded [E-204]. Assessment pending analyst confirmation.

All identifiers, times and values are mock data.

The operating loop

One loop from first signal to finished report.

Vanchi supports the whole intelligence cycle, not just detection. Select a stage.

Detect

Find meaningful events across multimodal maritime data — gaps, detections, encounters, deviations and registry changes.

Learn feeds Detect. Validated analyst feedback improves the next pass through the loop.

Capabilities

Four things Vanchi understands.

Understand entities

Who a vessel is — held as ranked, reversible hypotheses.

  • Vessel identity resolution
  • Ownership and operator networks
  • Flag and registry history
  • Confidence-aware association

Understand behaviour

What a vessel is doing — against its own baseline, not a global threshold.

  • Track and trajectory analysis
  • Encounters, loitering, AIS gaps
  • Port patterns and corridor deviations
  • Repeated behavioural signatures

Understand evidence

Why an assessment holds — and what would break it.

  • SAR, EO, VIIRS and RF association
  • Contradiction detection
  • Multimodal timelines and provenance
  • Calibrated confidence, alternative hypotheses

Support action

What to do next — with a human deciding.

  • Case prioritisation and next-best-action
  • Collection recommendations
  • Watchlists, briefings, reports
  • API and workflow integration

The workspace

A case moves through one surface.

Queue, map, evidence, hypotheses, tasking and report — connected, so nothing about a case is ever out of view.

Morning BriefWhat changed overnight, ranked by mission relevance.Feeds the queue
Multimodal TimelineEvery observation on one clock, per case.Opens from any evidence
Tasking WorkbenchCollection opportunities against open questions.Issues the SAR revisit
Report GeneratorStructured output, every claim linked to evidence.Closes the case

Architecture

Built for maritime reasoning.

Five layers. The model reasons and explains; deterministic services produce the numbers; analysts decide. Expand any layer.

ObserveMultimodal maritime data enters as evidence, never as assumed truth.
AIS and satellite AIS, SAR and EO detections, VIIRS, RF and radar-derived observations, registries, ownership records, port calls, sanctions and watchlists, weather, bathymetry, EEZ boundaries and mission holdings. Each observation carries source, time, uncertainty and access classification.
Core designSource list per deployment — confirm
ResolveEntity resolution, track association and spatiotemporal alignment.
Observations are normalised, aligned in space and time, and associated into physical tracks. Identity is kept separate from track and held as ranked, reversible hypotheses — two vessels are never merged simply because their broadcast identifiers match.
Core design
ReasonMultimodal inference over evidence: hypotheses, contradictions, confidence.
The foundation model reads the assembled evidence, compares hypotheses, detects contradictions, calibrates confidence and explains its assessment. Coordinates, distances and closest-approach values come from tested deterministic services the model calls — it does not improvise numbers.
Core design
ActPrioritisation, collection recommendations, alerts and reports.
Cases are ranked by mission relevance, risk and collection value. Vanchi recommends the next sensor, search or dataset; drafts evidence-linked reports; and routes every consequential conclusion to human review before it leaves the system.
Core designTasking-system integrations — confirm
LearnValidated analyst feedback improves future reasoning.
Analyst confirmations, corrections and rejections are captured as governed labels. Model updates are versioned and evaluated before release; historical assessments stay linked to the exact model version that produced them, and are never silently reinterpreted.
Core designContinual-learning cadence — planned

Evaluations

Evaluated for the work that matters.

Generic language-model benchmarks do not measure maritime judgement. Vanchi is evaluated on the analyst's actual tasks — against honest baselines, on held-out data split by time, geography and vessel identity.

Evaluation dimensionWhat it measuresResult
Vessel identity resolutionPairwise accuracy; false-merge and false-split rates[BENCHMARK RESULT]
Track associationPrecision and recall under gaps and clutter[BENCHMARK RESULT]
Dark-vessel candidate rankingRecall at fixed analyst workload[BENCHMARK RESULT]
Evidence-citation accuracyDo claims trace to the evidence cited[BENCHMARK RESULT]
Contradiction identificationConflicting sources surfaced, not resolved away[BENCHMARK RESULT]
Confidence calibrationStated confidence versus observed outcomes[BENCHMARK RESULT]
Collection-recommendation qualityDid the recommended collection resolve the question[BENCHMARK RESULT]
Report factual consistencyGenerated reports against adjudicated case files[BENCHMARK RESULT]
Performance under deceptionSpoofed identifiers, gaps and adversarial records[BENCHMARK RESULT]
Analyst time-to-assessmentMeasured against human-only baseline workflows[BENCHMARK RESULT]
vs. rule-based maritime analytics vs. single-source AIS analytics vs. generic multimodal models vs. human-only workflows vs. Vanchi ablations (per modality)

Results are published when evaluations complete. No figure appears on this page until it has been measured on held-out data and survived the deception suite. Analyst acceptance, correction and rejection rates are reported alongside accuracy.

View evaluation methodology

Evidence and trust

Every assessment should be inspectable.

Hover any underlined claim. In the product, this is how a reviewer audits a finding — down to the exact observation.

The vessel is assessed as the continuation of the gapped track, proceeding northeast at approximately 12 knots, with one contradicting observation held open, pending analyst confirmation after the next collection window.

E-201/2SAR + RF association, score components stored
SVC-CPADerived by deterministic track service, result ID cited
E-204VIIRS null return — recorded, weighted, not deleted
WF-118Human-review gate in case workflow, audit-logged
Source-level provenanceEvery claim carries its evidence identifiers, source and observation time.
Confidence and uncertaintyAssessments state calibrated confidence and what would change it.
Contradictions held openConflicting evidence is surfaced and retained, never resolved silently.
Human review by designConsequential conclusions route to authorised reviewers before release.
Audit historyAppend-only records of assessments, edits, dispositions and versions.
Access controlRole- and mission-based permissions with data-separation controls.
Configurable retentionRetention and deletion policies set per source and jurisdiction.
Sovereign pathwaysDeployment inside controlled environments, including air-gapped. [DEPLOYMENT OPTION — CONFIRM]
ObservabilityModel, rule and workflow versions visible on every output.

Vanchi supports evidence-based human decisions. It does not make enforcement, targeting or policy decisions, and has no actuation path to do so. Specific security certifications are stated only when attained.

Deployment

Three ways to run Vanchi.

Vanchi Workspace

The complete analyst environment — queue, cases, evidence, tasking and reporting — ready to operate.

[DEPLOYMENT OPTION — CONFIRM]

Vanchi API

Model, entity, event, evidence and workflow capabilities embedded into the systems your analysts already use.

[DEPLOYMENT OPTION — CONFIRM]

Sovereign Vanchi

Private-cloud, on-premises or air-gapped deployment for sensitive missions, with data remaining in your control.

[DEPLOYMENT OPTION — CONFIRM]

Integrations are scoped per deployment. Vanchi connects to the systems above through adapters; it does not claim universal compatibility.

Missions

Where Vanchi does its work.

Dark-vessel investigation
Difficult todayA gap is easy to detect and hard to explain. Analysts search imagery archives by hand, and most candidates are benign.
Evidence combinedAIS history, SAR and EO detections, VIIRS, RF observations, corridor priors, weather and reception baselines.
Decision supportedWhich gapped tracks deserve collection, which detections belong to them, and when the case is closed as benign.
Sanctions and deceptive-shipping analysis
Difficult todayIdentity laundering spans registries, flags and ownership shells. Each source is checked separately; contradictions slip through.
Evidence combinedIdentifier histories, ownership networks, flag changes, port calls, encounters and observed vessel characteristics.
Decision supportedWhether the vessel in front of you is the vessel of record — with the conflict trail attached for compliance review.
Illegal fishing detection
Difficult todayFleets go dark at zone boundaries. Enforcement assets are scarce and every patrol hour spent on a false positive is lost.
Evidence combinedVMS and AIS behaviour, zone geometry, activity patterns, SAR detections and licence records.
Decision supportedWhich incursions to task a patrol against, ranked by confidence and the cost of being wrong.
Critical infrastructure protection
Difficult todayLoitering near cables and platforms is visible only in hindsight, after the pattern is assembled manually.
Evidence combinedInfrastructure geometry, track behaviour, loitering and approach patterns, vessel identity and operator history.
Decision supportedWhich approaches warrant escalation now, and which vessels show a repeated pattern worth a standing watch.
Search and rescue support
Difficult todayThe first hour is spent reconstructing where the vessel was and who is close enough to help.
Evidence combinedLast known positions, drift and weather conditions, nearby traffic and distress communications.
Decision supportedWhere to search first and which vessels to task for assistance — with the reasoning available to the coordinator.
Shipping and insurance risk
Difficult todayUnderwriting and claims hinge on vessel history that is fragmented across providers and often self-reported.
Evidence combinedBehavioural history, encounters, gap patterns, ownership changes, port-state records and observed activity.
Decision supportedWhether the risk story a vessel tells matches the evidence trail it left — before the policy is written.

Research and documentation

Research behind Vanchi.

The technical record: how the model works, how it is evaluated, and how it should be deployed responsibly.

Model overviewArchitecture, modalities, reasoning approach and limitations.[PUBLICATION PENDING]
Evaluation methodologyThe maritime evaluation framework, baselines and deception suite.[PUBLICATION PENDING]
Evidence and provenance architectureHow claims stay linked to observations across the case lifecycle.[PUBLICATION PENDING]
Multimodal track associationAssociating SAR, EO, RF and cooperative tracks under uncertainty.[PUBLICATION PENDING]
Maritime ontologyEntities, events, evidence types and their relationships.[PUBLICATION PENDING]
Responsible deploymentHuman review, non-goals, access control and audit requirements.[PUBLICATION PENDING]
System cardIntended use, evaluation summary and known failure modes.[PUBLICATION PENDING]
API documentationModel, entity, event and workflow interfaces.[PUBLICATION PENDING]
Spatiotemporal reasoningHow the model handles time, geometry and deterministic services.[PUBLICATION PENDING]

Working with Vanchi

Understand the ocean as a system.

Bring multimodal maritime evidence, model reasoning and analyst judgement into one operational loop.

Request Access Speak with the Vanchi Team