Your fleet has a documentation blind spot.
We close it.
Cited answers. Verified citations. Transparent confidence.
Onboard every vessel.
Works when the link drops · SMS data stays on the ship
Chatbots hallucinate. RAG approximates.
We retrieve, cite, and refuse when we can't.
- 1.Mark position — activate MOB waypoint on GPS[1]
- 2.Throw lifebuoy with light and smoke marker[1]
- 3.Post lookouts — minimum two, continuous visual[1]
- 4.Sound general alarm — three long blasts[1]
- 5.MAYDAY / PAN-PAN on VHF 16 if required[1,2]
- 6.Launch rescue boat with trained crew[1,2]
- 7.Notify Coast Guard and nearby vessels[2]
- 8.Continue search until located or officially suspended[2]
- 9.Headcount and muster — verify all crew[1]
Running onboard. No internet required.
The ISM gap
The SMS only works if the crew can find the right procedure in time.
4,103
ships detained by Paris MoU Port State Control in 2024 — a 4.03% detention rate, up from 3.81% the year before.
Source · Paris MoU 2024
17.2%
of Paris MoU deficiencies in 2024 fell in fire safety — the single largest category. ISM-related deficiencies appear in 4.6% of cases.
Source · Paris MoU 2024
77,526
deficiencies recorded by Tokyo MoU across 26,140 inspections and 1,189 detentions in 2024. 54% fell in just four categories.
Source · Tokyo MoU 2024
70–78%
of maritime accidents analysed in peer-reviewed HFACS studies of MAIB, TSB and NTSB casework are attributed to human factors — with a recognised class characterised by non-compliance with the SMS.
Source · MDPI 2023 HFACS analysis · Taylor & Francis 2025
“The procedure existed. The crew could not find it in time.”
That is an ISM finding, not a technology problem — and it's what we help you close.
The problem
A single vessel carries 1 pages.The crew sees one at a time.
Fire Response
Bridge Emergency
SOLAS III — LSA
Muster List
COLREGs Rule 19
Enclosed Space Entry
Stability Booklet
Man Overboard
Bridge Team Mgmt
Anchor Watch
Full Fathom sees every page at once — and ranks SMS above regulation, vessel above reference, by design.
Four steps to operational safety intelligence
Upload
Upload your SMS manuals, procedures, and regulatory references. We parse, chunk, and build an AI knowledge index — preserving tables, numbered procedures, and cross-references.
Deploy
Download a 5GB bundle to each vessel. Runs on any mini-PC. Zero internet required. Crew access it through a browser on any device connected to the vessel network.
Trust
Every answer shows its source tier (SMS, vessel, regulation, reference), a High / Medium / Low confidence badge, and whether each citation is verified against what was actually retrieved. Crew see at a glance whether to act or verify.
Learn
See what your crews are asking. Find documentation gaps. Identify procedure hotspots and confusion signals. Fix them before the next audit.
Upload
Upload your SMS manuals, procedures, and regulatory references. We parse, chunk, and build an AI knowledge index — preserving tables, numbered procedures, and cross-references.
Deploy
Download a 5GB bundle to each vessel. Runs on any mini-PC. Zero internet required. Crew access it through a browser on any device connected to the vessel network.
Trust
Every answer shows its source tier (SMS, vessel, regulation, reference), a High / Medium / Low confidence badge, and whether each citation is verified against what was actually retrieved. Crew see at a glance whether to act or verify.
Learn
See what your crews are asking. Find documentation gaps. Identify procedure hotspots and confusion signals. Fix them before the next audit.
Same prompt, three systems. One survives scrutiny.
In a fire, one missed step is a life. Only grounded answers belong at sea.
The pipeline
Fifteen stages. Every question.
No single model answers your crew. Every question passes through an engineered pipeline before a word reaches the bridge.
Question Sanitisation
Production-hardened input handling
Strips typographic characters that would break downstream processing.
Query Decomposition
Compound questions split into retrieval paths
Two-part questions become two searches whose results are merged.
Domain Vocabulary Expansion
Bridges crew language to regulatory text
Twenty curated maritime mappings so phrasing never blocks the right document.
Multi-Signal Retrieval
Semantic + keyword, mathematically fused
Two independent searches converge via Reciprocal Rank Fusion.
Relevance Feedback Expansion
Second-pass retrieval at zero LLM cost
The system searches again using the best first-pass results as seeds.
Neural Reranking
Cross-attention precision scoring
Twenty candidates re-scored by a dedicated model; the top three rise.
Authority Hierarchy
SMS outranks regulation, by design
Tier multipliers nudge rerank scores so your procedures lead and regulations corroborate.
Confidence Gating
Multi-threshold quality control
Scores below hard threshold are discarded; below soft, the model is told to flag uncertainty.
Self-Correcting Search
Detects its own failures and retries
Rewrites in formal regulatory vocabulary and runs a second retrieval pass.
Intelligent Context Assembly
Parent-child expansion, diversity enforced
Retrieved chunks pull in their parent section so procedures never arrive truncated.
Attention-Optimised Ordering
Matching how language models actually read
Most relevant material placed at start and end of context -- the high-attention zones.
Cited Answer Generation
Every claim traced to its source
Inline numeric citations generated with the answer, resolved to doc/section/page.
Citation Validation
The hallucination guard
Every citation number is checked against the chunks that were actually retrieved.
Sufficient-Context Autorater
Based on Google ICLR 2025 research
A second model checks whether the context actually contained enough to answer.
Confidence Scoring
High / Medium / Low with transparent reasoning
Four signals combine into the badge the crew sees. Refusals are length-aware.
Question Sanitisation
Production-hardened input handling
Strips typographic characters that would break downstream processing.
Query Decomposition
Compound questions split into retrieval paths
Two-part questions become two searches whose results are merged.
Domain Vocabulary Expansion
Bridges crew language to regulatory text
Twenty curated maritime mappings so phrasing never blocks the right document.
Multi-Signal Retrieval
Semantic + keyword, mathematically fused
Two independent searches converge via Reciprocal Rank Fusion.
Relevance Feedback Expansion
Second-pass retrieval at zero LLM cost
The system searches again using the best first-pass results as seeds.
Neural Reranking
Cross-attention precision scoring
Twenty candidates re-scored by a dedicated model; the top three rise.
Authority Hierarchy
SMS outranks regulation, by design
Tier multipliers nudge rerank scores so your procedures lead and regulations corroborate.
Confidence Gating
Multi-threshold quality control
Scores below hard threshold are discarded; below soft, the model is told to flag uncertainty.
Self-Correcting Search
Detects its own failures and retries
Rewrites in formal regulatory vocabulary and runs a second retrieval pass.
Intelligent Context Assembly
Parent-child expansion, diversity enforced
Retrieved chunks pull in their parent section so procedures never arrive truncated.
Attention-Optimised Ordering
Matching how language models actually read
Most relevant material placed at start and end of context -- the high-attention zones.
Cited Answer Generation
Every claim traced to its source
Inline numeric citations generated with the answer, resolved to doc/section/page.
Citation Validation
The hallucination guard
Every citation number is checked against the chunks that were actually retrieved.
Sufficient-Context Autorater
Based on Google ICLR 2025 research
A second model checks whether the context actually contained enough to answer.
Confidence Scoring
High / Medium / Low with transparent reasoning
Four signals combine into the badge the crew sees. Refusals are length-aware.
All fifteen stages run per question — offline, on vessel hardware. 290 tests gate every build.
What we stand behind
Four things we'll commit to in writing.
Two quantitative facts about the runtime itself. Two qualitative facts about how it behaves. The numbers you really want — latency, retrieval quality, corpus performance — we measure per build on the reference rig and publish under pilot.
engineered stages per answer
Sanitise · decompose · retrieve · rerank · rank by authority · gate · self-correct · assemble · order · generate · validate · autorate · score.
runtime binary, one file
The onboard server. Statically linked. Cold-start under two seconds on the reference build. No internet required.
citation validated before the crew sees it
If the answer cites a source, the validator checks it against the chunks actually retrieved. If it does not pass, the crew sees a warning — never a confident fabrication.
when the context is insufficient
A Google ICLR 2025-style sufficient-context autorater inspects the retrieved context independent of the generated answer. If the context cannot support the question, the system says so.
Internal measurements refer to the current reference build on the Propel Group engineering rig. They are representative, not contracted. Your fleet measurements will differ; we size hardware as part of the pilot.
The artefact
27 megabytes. One file. One ship.
Everything you have seen so far — the retrieval stack, the authority hierarchy, the citation validation, the confidence scoring — ships as one statically-linked binary. This is the real startup sequence, with real latencies, from the runtime on a vessel.
Category comparison
Why Full Fathom vs. general AI assistants
The maritime safety stakes are different. The requirements on the AI have to be different too.
- Generic chatbot
- No
- Enterprise RAG
- Document-level
- Full Fathom AI
- Chunk-level with page numbers
- Generic chatbot
- —
- Enterprise RAG
- —
- Full Fathom AI
- Yes — flags unverified
- Generic chatbot
- —
- Enterprise RAG
- —
- Full Fathom AI
- High / Medium / Low with reasoning
- Generic chatbot
- —
- Enterprise RAG
- —
- Full Fathom AI
- Vessel / SMS / Regulation / Reference
- Generic chatbot
- Summarises
- Enterprise RAG
- Summarises
- Full Fathom AI
- Full mandatory procedure enforced
- Generic chatbot
- —
- Enterprise RAG
- Sometimes
- Full Fathom AI
- Always, with transparent rationale
- Generic chatbot
- No
- Enterprise RAG
- Varies
- Full Fathom AI
- Fully offline by design
- Generic chatbot
- No
- Enterprise RAG
- Yes
- Full Fathom AI
- Yes
- Generic chatbot
- No
- Enterprise RAG
- No
- Full Fathom AI
- Yes — SMS outranks regulation
Built for three roles
The crew asks. The DPA audits. The fleet learns.
For crew onboard
- Ask in plain English — no training required
- Every answer cites the source document, section, and page
- Works when the link drops; runs on the vessel network
- Accessible from any browser — desktop, tablet, phone
For the DPA / HSEQ team
- Prove your SMS is followed with a query + citation log per vessel
- Map every answer to an ISM Code element and a PSC category
- Authority hierarchy ensures your SMS leads, regulation corroborates
- Refuses rather than guesses when context is insufficient
For fleet managers & superintendents
- Deploy once, update on port calls
- See what crews ask, and what the system couldn't find
- Documentation gaps surface as unanswered questions
- Filter by vessel, vessel type, and time period
Built for safety-critical operations
Fully Offline
All inference happens onboard. No cloud processing of queries.
No Crew PII
Query logs contain no personal data. Zero crew identification.
Verified Citations
Every citation is checked against what was actually retrieved. Invented or mismatched citations are flagged with a warning — not shown as facts.
Confidence Transparency
Every answer shows High, Medium, or Low confidence with a tooltip explaining why. Crew know when to trust and when to verify.
Authority Hierarchy
Your SMS procedures outrank generic regulation by design. Vessel > SMS > Fleet > Regulation > Reference, with every source tagged.
Safety-Complete Procedures
Man overboard, fire, enclosed space: we return every mandatory step, not a summary. Completeness enforced by design, not hoped for.
ISM Aligned
Supports ISM Code compliance evidence across Sections 6, 7, and 12.
Tamper-Evident Bundles
SHA-256 checksums on every bundle. The runtime refuses to load a file that has been altered in transit.
Your Data, Isolated
Full multi-tenant data separation. No cross-tenant access.
Hardware requirements
Runs on hardware most vessels already have. No GPU required.
| Minimum | Recommended | |
|---|---|---|
| CPU | x86_64, 4 cores | x86_64, 8 cores |
| RAM | 16 GB | 32 GB |
| Storage | 100 GB SSD | 256 GB SSD |
| GPU | Not required | Not required |
| Network | Vessel LAN | Vessel LAN |
One Fleet plan. Per vessel, annual.
USD, annual billing, volume bands by fleet size. Setup fee scales with Band A and is waived for design partners. Premium features priced per fleet per month.
— The deep end —
Bring certainty onboard. Prove it ashore.
Two months, two vessels, no card. Your SMS ingested, your crew using the real product onboard. Decide at the end of month two.