Ingested ashore. Answered onboard.
A safety-critical reasoning system for maritime. Crew get a complete procedure, a source tier, and a confidence score on every answer — onboard every vessel, offline.
From documents to deployable knowledge.
Upload PDFs, manuals, procedures, and regulatory references through the web dashboard. The platform parses, chunks, embeds, and packages everything into a single versioned bundle — every document in your fleet, searchable in one place.
- SMS manuals, vessel-specific procedures, equipment guides, SOLAS / MARPOL references
- Parent-child chunking preserves numbered procedures and tables
- 1,024-dimension semantic index plus full-text keyword index
- Versioned, SHA-256 checksummed, ~5–6 GB typical bundle
Fifteen stages. Every question.
No single model answers your crew. Every question passes through an engineered pipeline — sanitisation, decomposition, multi-signal retrieval, neural reranking, authority hierarchy, self-correcting search, intelligent assembly, cited generation, and a hallucination guard.
- Hybrid semantic + keyword retrieval with Reciprocal Rank Fusion
- Cross-attention reranking surfaces the top three sources from twenty candidates
- SMS outranks vessel, vessel outranks regulation — by design
- Self-correcting retry when context is thin; every claim traced to source
Every answer carries its source tier.
Crew see the ruling, the source document (with section and page), and a confidence score derived from reranker, diversity, citation verification, and self-correction signals. A warning appears instead of an answer when context is insufficient — never a guess.
- High / Medium / Low confidence with transparent reasoning
- Citation validation checks every bracketed number against retrieved chunks
- Length-aware refusal detection — short 'not in the bundle' classed correctly
- Logged for shore-side knowledge-gap analytics, cached semantically on repeat
A number that earns itself.
The confidence score isn't a single model output — it's a running total built from distinct signals as the pipeline runs. You see each delta accrue; you see where the score came from; you see the final number locked.
- Retrieval relevance + reranker margin contribute the first half
- Citation verification (every bracketed [n] bound to a chunk) contributes the next
- Sufficient-context autorater and self-correction close the final gap
- If the running total ends below threshold, the system refuses rather than guesses
Fleet-synchronised. Zero-downtime rollback.
When you update a procedure shore-side, the platform rebuilds the bundle and the vessel receives it on its next connectivity window. VSAT at sea, broadband in port, USB if neither. Every bundle is checksummed end-to-end; if verification fails, the update is rejected. Failed updates roll back automatically.
- Versioned bundles — you always know what every vessel is running
- SHA-256 verification before apply, rollback on any failure
- Delivery over VSAT, broadband, or sneakernet as the route demands
- Query logs (no crew PII) flow shore-side for fleet-level gap analysis
The view from the bridge. Cited, immediate, offline.
One browser tab on any device on the vessel network. Crew ask in plain English; every answer shows the source document, section, and page. No training. No internet. No drill-down through folders of PDFs when the alarm is ringing.
Crew type the question they'd ask a senior officer. No keywords, no syntax, no folder diving.
Every response shows the source tier, the document, the section, and the page. Never a bare paraphrase.
Runs on a single mini-PC on the vessel network. No internet during a query. Ever.
Shown · onboard runtime on a reference vessel. Real deployments brand with your fleet's name and vessel identifier.
Every answer, ashore. At a glance, on a chart.
Every crew query on every vessel rolls up to the shore dashboard. Volume per day, refusal rate, response time, and the topics crews are actually asking about. Knowledge gaps surface as the questions the system couldn't answer, directly from the data — before they become audit findings.
30-day rolling history per vessel or fleet-wide. Weekly rhythm and trend visible at a glance.
Top unanswered topics surface automatically. Each one is a documentation hole worth closing.
Crews' actual concerns grouped and ranked. The procedures that matter show themselves.
Shown · representative dashboard with a reference 12-vessel fleet. Live tenants see their own query data, anonymous at the fleet aggregate level.
— Ready when you are —
Close the documentation gap. Onboard.
Two months, two vessels, no card. Your SMS ingested, your crew using the real product onboard. Decide at the end of month two.