Maritime Risk Intelligence
MARITIME RISK INTELLIGENCE PLATFORM

From inspection question
to charter decision — consistently,
auditably, on your network.

A QA standardisation system for vessel vetting. Repeatable scoring, cited evidence, controlled escalation — running entirely on your infrastructure.

Founder Briefing
2026 · QA Team Briefing
01FRAMING
THE OBJECTION YOU’LL HEAR

“Isn’t this just an LLM wrapper?”

What a wrapper can do

01Read a PDF and summarise it
02Score findings with a prompt
03Look impressive in a single demo
04Disagree with itself on the same input

What a wrapper cannot do

→Run on-premises with zero outbound traffic
→Return the same answer twice on the same input
→Cite page, line, and quote for every claim
→Escalate low-confidence calls to a human reviewer — when the system disagrees with itself on the same input, the case goes to a human for clarification
We agree the LLM is one piece. The platform is everything around it.
02TODAY

A vetting decision today — by hand.

Every charter starts with a senior surveyor reading a stack of inspection PDFs, cross-checking certificates, and forming a judgement that doesn’t transfer to anyone else.

01 · SCALE
0
documents per vessel: SIRE, PSC, Q88, HVPQ, P&I, class certificates…
02 · TIME
0
to read, cross-reference, score and write up — per ship.
03 · KNOWLEDGE
0
before a junior analyst can decide alone — senior judgement doesn’t transfer.
60+ PDFs → Senior surveyor → 60 min → One charter decision
QA managers don’t need an AI that reads faster. They need one that decides repeatably.
03FOUNDATION

Every finding is one triple.

Q 9.4Is the inert-gas system fitted with a fixed oxygen analyser?
OSystem fitted, fixed oxygen analyser inoperative since last drydock.
REquipment critical to fire prevention is non-functional.HIGH

Every High / Medium / Low / Positive in the report is built from exactly this triple — no more, no less.

04ARCHITECTURE

How a finding flows through the platform.

Five engineered stages. One model call. Verbatim extraction is the work.
05AORCHESTRATION
PHASE 1 · SPAWN

The Master Agent wakes — six specialists in parallel.

PSCrunning · ~280ms
P&Irunning · ~310ms
SIRErunning · ~640ms
CLASSrunning · ~210ms
SANCTIONSrunning · ~150ms
ITFrunning · ~190ms
t = 0ms  ·  spawn(specialists, parallel)  ·  each agent receives only the documents it needs.
05BORCHESTRATION
PHASE 2 · REPORT · VOTE

Specialists return. Five votes collapse to one answer.

SELF-CONSISTENCY VOTE
RUN 1HIGH
RUN 2HIGH
RUN 3MED
RUN 4HIGH
RUN 5HIGH
mode → RISK = HIGH
disagreement > 1/5 → analyst queue
no decision is forced from noise.
fuse(findings, charter_weights)  ·  Master Agent analyzes specialist outputs and writes the final charter-weighted score.
05CKNOWLEDGE BASE
PHASE 3 · LEARN · ACCELERATE

The platform learns once. Every future inspection runs faster.

INSPECTION 001
47.0s
  • 1Static vessel information is stored once. Ship history, prior incidents, and known vessel characteristics are retained — no repeated research needed.
  • 2Same inputs give the same answers. If the platform has seen the finding before, the answer is retrieved instantly — no LLM call needed.
  • 3Knowledge stays inside your environment. The shared knowledge base exists within your company scope. Your data never leaves.
  • 4Future inspections get faster. Prior ship research does not need to be repeated — knowledge compounds with every inspection.
The AI handles only what it has never seen before.  ·  everything else is recalled from your company’s knowledge base.
08APERSONIFICATION
ADAPTIVE INTELLIGENCE

Every QA becomes a digital twin. The platform learns how they think.

  • 1Each QA's writing style and evaluation logic is captured and learned on every use.
  • 2Company profile configured at deploy. Charter weights, risk appetite, terminology preferences baked in.
  • 3Agents critique each other before final response. Orchestrator resolves disagreement; one answer reaches the QA.
  • 4Non-sensitive patterns flow up to the shared KB. Every user benefits as the platform matures across all companies.
Personalization compounds.  ·  Day 1 is calibration. Day 90 is a colleague.
06ACCURACY
REPEATABILITY · NOT MAGIC

Same inputs. Same answers.

We engineer for consistency where we can. We escalate where we can’t.

Extraction is the work. Repeatability comes from the cache, not the model.
07WEIGHTING

Charter-type weighting. Industry guidance + per-customer delta.

SIRE
30%
PSC-correlation weighted · Paris MOU
PSC
20%
INHERENT
30%
CLASS
20%
INHERENT = HVPQ + Q88 + PIQ — the owner’s side
SANCTIONS hard-stop overrides any score · positives reduce risk
OUTPUT · CHARTER DECISION
GO  ·  weighted score ≤ threshold
REFERRAL  ·  analyst review
NO-GO  ·  sanction or hard-fail
modeled on the Adnoc 3-state pattern.
each customer can tune their delta on top.
Industry guidance + your house rules = your decision boundary, not ours.
07BSIRE SCORING
IDWAL-INSPIRED METHODOLOGY

SIRE score weighted by detention probability. Not severity alone.

BEFORE — simple severity tiers
sire_raw =
(HIGH × 15) +
(MED × 5) +
(LOW × 1.5)
AFTER — PSC-correlated
sire_raw = Σ (
  finding
  × severity_weight
  × psc_correlation_weight
)
Same findings. Each one now carries its
historical likelihood of causing detention.
DETENTION CORRELATION BY CATEGORY
Fire Safety
1.8×
ISM / SMS
1.6×
Life-Saving Appliances
1.5×
Navigation / ECDIS
1.4×
Documentation
0.8×
Crew Welfare
0.7×
Source: Paris MOU Annual Report  ·  public · free · defensible to any oil major.
07AREFERRAL
CONTROLLED ESCALATION

Referral is not a rejection. It’s a structured loop.

When the system can’t decide cleanly, it routes to QA — with the gaps already flagged.

WHEN REFERRAL FIRES

  • ·Conflicting evidence across documents
  • ·Partial rectification of prior findings
  • ·Recurring low-grade observations — systemic pattern
  • ·PSC vs. SIRE signals contradict
  • ·Operator-specific tolerance gap
  • ·Missing supporting records or evidence

RE-EVALUATION LOOP

  • 1AI flags REFERRAL with rationale + evidence gaps
  • 2QA team investigates · uploads additional records
  • 3AI re-evaluates against the updated dataset
  • 4Final outcome reissued: GO · REFERRAL · NO-GO
We empower QA teams by making decades of inspection knowledge available from day one.
07BRISK METHODOLOGY
IN-HOUSE RISK METHODOLOGY

Risk is not decided by AI. It follows predefined rules.

The system consults our in-house risk methodology — AI identifies findings, rules determine the risk level.

RISK CLASSIFICATION

HIGH
Immediate operational or safety threat
fatality onboard · pollution incident · sanctions hit · critical equipment inoperative · ISM major non-conformity
MEDIUM
Elevated concern requiring review
recurring PSC observations · partial rectification · overdue certification · flag-state risk pattern
LOW
No significant concern identified
minor observations · rectified deficiencies · positive crew and management record

SCORING LOGIC

Overall score is calculated from sub-scores across document categories. Risk criticality adjusts the score based on finding severity.
retention finding → score reduced
high-risk finding → score reduced further
positive finding → score offset
SANCTIONS → hard stop · overrides any score
RULE, NOT RANDOM
The AI consults the methodology. It does not invent risk levels. Each score drop follows a predefined business rule.
Risk classification and scoring are separate. How a finding becomes high-risk is distinct from how the overall score is calculated.
08CONFIDENTIALITY

Your data never leaves your network.

Public cloud
Off-network. Out of scope. Not permitted.
Customer GPU
In your VPC. Behind your firewall. Yours.
  • 1Open-weight 70B+ models — Llama 3, Qwen 2.5, DeepSeek class — running on customer GPU.
  • 2Air-gapped deployment available. Model + classifier ship as a Docker image with zero outbound network.
  • 3Frontier models for R&D · open-weight for production. The production deployment is yours, in your VPC, behind your firewall.
Default GPU footprint: 2× H100 / 8× A100 per region.
08AINTEGRATION
PLUG-IN, NOT RIP-AND-REPLACE

It fits your workflow. Not the other way around.

YOUR EXISTING SYSTEMS
Q88 / HVPQ records
SIRE 2.0 reports
PSC feeds
P&I records
Sanctions watchlists
Internal vessel database
→
MASTER AGENT + KB in your environment
→
YOUR EXISTING WORKFLOW
QA dashboard
Charter desk handoff
Audit log
Analyst review queue
Connect what you already have. No process change. No new tools to learn.  ·  review time drops from ~60 minutes to ~1 minute.
09RELIABILITY

Reliability is layered, not promised.

Page 14 → “fixed O₂ analyser inoperative” → Q9.4 → severity:HIGH → cite:0x4a2f…
We don’t ask the customer to trust the AI. We ask them to audit it.
10CO-PILOT

We empower QA teams by standardizing decision-making from day one.

WITHOUT THE PLATFORM

60 min

per report. 60+ documents. Senior surveyor, irreplaceable. New hires take six months to ramp — and the senior’s reasoning is locked in their head.

→

WITH THE PLATFORM

<60 sec

per report. One dashboard. Junior analyst is effective on day one because the senior’s reasoning is now captured, queryable, and audit-able.

+ provenance + override + explanation + cited quotes + knowledge transfer
New analyst → senior reasoning, day one. Senior analyst → freed for the calls only humans should make.
10AASK
INTERACTIVE QA COPILOT

Beyond scoring. Analysts can probe deeper.

SAMPLE QUESTIONS
Why was this vessel referred instead of approved?
Show recurring IG observations across the last 3 SIRE inspections.
Compare PSC deficiencies against this operator’s trend.
Has this operator had repeated detentions fleetwide?
RULES
✓Retrieves only from approved internal sources
✓No public internet access
✓Every answer cited · overridable by an analyst
✓Identical question on identical data → identical answer
✗Not a chatbot. Not the public internet.
QA intelligence layer, not an AI assistant. Faster, defensible decisions.
11OUTPUT

What the QA team actually sees.

VESSEL
[ redacted ] · IMO ********
M MEDIUM
2
High
5
Med
11
Pos
DECISION · REFERRAL
ANALYST QUEUE · 1 disagreement
SIRE Q9.4 · HIGH
Inert-gas system · fixed O₂ analyser inoperative.
“System fitted but fixed oxygen analyser inoperative since last drydock.”
page 14 · cite:0x4a2f… override
PSC · 2024-03 · MEDIUM
Detained Rotterdam · deficiency rectified at port.
“Vessel detained 2024-03-11 · ISM non-conformity · released 2024-03-13.”
PSC report · cite:0x9ab4… override
SIRE 6.1 · POSITIVE
Crew competence rated above benchmark.
“Crew demonstrated familiarity with cargo operations and emergency procedures.”
page 22 · cite:0x7d12… override
12DELIVERY

Where we are. Where we’re going.

TODAY MVP

  • ✓SIRE 2.0 inspection report end-to-end
  • ✓Verbatim quote provenance per finding
  • ✓Frontier models for R&D · open-weight for production
  • ✓Web UI demo with override + audit trail
  • ✓Validated against client-labelled cases

PRODUCTION PHASE 2 · 90 DAYS

  • →All 6 specialist agents online
  • →Multi-document fusion (SIRE + PSC + P&I + Class + Sanctions + ITF)
  • →On-premises open-weight deployment
  • →Hash-cache + self-consistency layer live
  • →Audit trail + analyst review queue
  • →SOC 2-aligned operational controls
13ROADMAP

Phased delivery.

WEEK 0
DEMO ✓
already shipped
1 vessel · SIRE end-to-end · web UI
WEEK 2
METHODOLOGY DECK ✓
this document
20-slide QA team briefing
WEEK 8
PILOT
one vessel class · one customer · on-prem · all 6 agents
target latency <90s · audit trail v1 · analyst queue
WEEK 24
PRODUCTION
multi-customer · audit-grade · SOC 2 path
multi-customer · cached findings replay deterministically
WEEK 32
BILINGUAL
EN / JA toggle · QA copilot responds in Japanese
UI · chatbot · report output fully bilingual
60-day pilot, 6-month production. Honest commitments, not aspirational ones.
14CLOSE
“We empower QA teams by helping
standardize decision-making.”
hello@…
Maritime Risk Intelligence
01/17