"MarineGPT is not a chatbot. It is a maritime decision system — a 7-stage analytical pipeline backed by 14 computation engines, delivering institutional-grade outputs for ship owners, investment committees, and maritime professionals worldwide."
14 Computation Engines
Every answer is computed, not generated. Here is what runs before the response reaches you.
Gate 0 — runs BEFORE all engines. 10 hard-stop rules: sanctions jurisdiction, evasion signals, criminal activity, prohibited cargo, negative voyage days, impossible vessel speed, terminal unviability. HARD_STOP blocks all downstream computation.
HARD_STOP • CONDITIONAL • WARNINGTCE, break-even freight, voyage profit, bunker sensitivity. TCE_DIRECT path when TCE is authoritative — skips fuel/port recomputation. Conflict detection when both TCE and full voyage inputs are present.
TCE = (Revenue − Port − Fuel) ÷ Voyage DaysAttained gCO₂/dwt·nm → Rating A–E per IMO MEPC.339. Emission factors per fuel type, corrective action thresholds, trajectory projections.
CII Rating A → E | EMISSION_FACTORS mapDispute expected value with maritime-calibrated base rates. Quantifies litigation value before commit.
EV = P(win)×(Claim−Cost) − P(lose)×CostQuantifies off-hire exposure, performance claims, laytime disputes, and deviation risk from charter party terms.
Off-hire • Performance • Laytime • DeviationStructured Bull / Base / Bear scenario generation with breakeven analysis and probability weighting across all three states.
Bull / Base / Bear | breakeven + P-weightsTechnical failure diagnosis with SOLAS implications, repair cost estimation, and classification society survey triggers.
SOLAS implications | repair estimate4-domain risk matrix covering Technical, Legal, Financial, and Regulatory exposure with severity scores and mitigation actions.
Technical / Legal / Financial / RegulatoryStress tests the base case against shock parameters. Classifies outcome as ROBUST, CONDITIONAL, or FRAGILE. Identifies breakpoint thresholds.
ROBUST / CONDITIONAL / FRAGILEPD from DSCR mapping (14 breakpoints), Altman Z-Score maritime adaptation (Z' for private companies), LGD by contract type, composite PD (65% worst-case + 35% average). HARD_STOP at PD ≥ 15%.
EL = PD × LGD × EAD | CVA | Z' ScoreScore 0–100 from firmness (40), broker agreement (25), recency (25), rate range (10). Validates 29 broker patterns. Deduplicates sources. Time-weighted freshness decay.
Score 0–100 | 29 broker patternsNewton-Raphson IRR solver. Hold NPV across Bull/Base/Bear scenarios. Break-even TCE via iteration. Scrap value: DWT × LDT/DWT ratio × $500/LDT.
IRR (Newton-Raphson) | HOLD / SELL / REDUCE5-year avg TCE benchmarks per segment, orderbook % by vessel class, structural score 0–100. Cycle phase classification across five states.
PEAK / PLATEAU / DECLINE / TROUGH / RECOVERYFull Black-Scholes implementation (d1/d2/N(d1)/N(d2)). Volatility benchmarks per vessel type. Computes option value of waiting vs fixing now.
FIX_NOW / WAIT / CONDITIONAL_FIX / PARTIAL_FIXThe Journey
Maritime Professionals Drowning in Data, Starving for Decisions
The observation that started everything: Masters, charterers, and ship owners were surrounded by data — AIS feeds, weather routing, freight indices, charter party clauses — but had no system that could synthesise it into a decision. Generic AI assistants failed because they had no concept of TCE, no SOLAS context, no maritime base rates. The industry was operating on instinct where it should be operating on intelligence.
The problem was never data. It was always the absence of a reasoning system that understood the sea.
The 7-Stage Pipeline Design — Not a Chatbot
The core architectural decision: MarineGPT would not be a chatbot with maritime data plugged in. It would be a 7-stage analytical pipeline — Situation Definition, Input Extraction, Deterministic Calculations, Root Cause Diagnosis, Scenario Comparison, Risk Mapping, Decision Directive — where every stage produces a structured output consumed by the next. The LLM sits at the end, articulating what the pipeline has already computed.
Every answer is computed, then communicated. The LLM does not guess — it narrates.
14 Computation Engines — Each Deterministic
Built sequentially, each engine deterministic: Voyage Economics (TCE, break-even), CII compliance, Legal EV, Charter Party Risk, Scenario generation, Root Cause diagnosis, Risk Mapping, and the Robustness stress tester. Every calculation uses tagged inputs (TOOL_SOURCED / ASSUMED / USER_PROVIDED) with explicit confidence levels. Monte Carlo simulation, Cornish-Fisher VaR, and Newton-Raphson IRR followed.
Tagged assumptions are not a UX feature. They are the difference between advice and accountability.
Binary Rejection Gate — Because Wrong Answers Cost Millions
The hardest-earned insight: before any engine runs, a gate must exist that can say no. The Binary Rejection Engine was built with 10 rules — sanctions jurisdiction, evasion signals, criminal activity, prohibited cargo, terminal unviability, impossible vessel speeds. A HARD_STOP blocks all downstream computation and returns a structured refusal with a legalPath — what the user CAN do instead.
A system that cannot say no is not a decision system. It is a liability.
Four Specialist Modes Serving Every Maritime Discipline
The platform expanded from a single decision engine into four specialist modes: Decision Engine (C-suite strategic advisor, 7-stage pipeline, 14 engines), Risk Analyst (institutional risk, Monte Carlo, Altman Z-Score, DSCR, Cornish-Fisher VaR), Operations Desk (navigators and operators, real-time tools, COLREGS, MARPOL playbook), and Maritime Academy (adaptive difficulty, 5 competency levels, concept graph). One platform. Every maritime discipline.
Different decisions carry different weight. The architecture had to match.
Impact By The Numbers
Our journey measured in real-world impact
Core Principles
Domain Expertise First
MarineGPT does not apply generic AI to maritime problems. It embeds maritime knowledge — SOLAS, MARPOL, charter party clauses, IMO regulations, freight indices — into the computation layer itself. The LLM executes reasoning grounded in domain fact.
Deterministic Over Generative
Every financial figure, every risk rating, every scenario output is computed by a deterministic engine before the language model sees it. The model narrates; it does not calculate. This eliminates hallucinated numbers from responses.
Professional Intelligence Standards
Outputs are held to institutional standards: tagged assumptions with confidence levels, Altman Z-Score credit ratings, DSCR-to-PD mapping, Cornish-Fisher VaR for tail risk, Newton-Raphson IRR for capital decisions. The bar is an investment committee, not a FAQ page.
Built for Weight-Bearing Decisions
Every architectural choice was made with the assumption that the decision on the other end carries real financial, legal, or safety consequence. This is why the Binary Rejection Gate exists. This is why assumptions are tagged. This is why HARD_STOP blocks downstream computation.