MarineIT / Platform / How the AI works
Parse · Enrich · Act
What it does, what it will not do, and where it fails.
The principle
Nothing on a screen without its workings.
A number your desk cannot check is a number your desk will quietly redo in a spreadsheet — which is exactly the behaviour the system was bought to stop. Worse, at some point somebody will stop checking, and a wrong figure will reach a recap.
So the rule the product is held to is narrow and absolute: if we cannot show you where a figure came from, it does not go on a screen. Every extracted value keeps its source text. Every computed figure opens to its inputs. Every override is recorded against the original.
An estimate you cannot audit is an estimate you cannot defend.
What we are not claiming
It does not fix cargoes. No agent negotiates on your behalf, and none will. Chartering is a judgement business conducted between people who know each other.
It does not predict the market. We benchmark against assessments parsed from your own market reports. We do not forecast where freight is going and do not sell a view on it.
It does not replace anyone. It removes transcription and triage. The decisions it surfaces still need a person, which is why every one of them opens.
It is not always right. Parsing fails, sources disagree, assumptions turn out badly. What matters is whether the system tells you.
Three stages
What happens between a circular arriving and a number appearing.
Each stage below states what it does and, more usefully, what it deliberately does not.
Read the message
Telex-format circulars, PDF recaps, market reports and Teams messages are read on arrival. Vessel particulars, cargo, quantity and tolerance, ports, laycan, rates, laytime terms and commission are extracted into structured fields, each keeping the exact text it came from.
Does not guess at a field it cannot read, discard a message it cannot parse, or normalise away a difference between two brokers’ wording.
Check it against everything else
Particulars from a circular are compared against the vessel’s own record and against live data — Lloyd’s for particulars, Netpas for distances, Tideform for bunkers, StormGlass for weather. Where sources disagree, both are kept and the discrepancy shown.
Does not average conflicting sources into one number, prefer one silently, or learn from your data.
Estimate, benchmark, check
TCE and freight estimated from the enriched inputs and your own assumptions. Benchmarked against the nearest Baltic route from your parsed market reports. Loadability checked against port draft, TCP and stowage factor.
Does not send an offer, accept a counter, commit you to anything, or hide an assumption inside a total.
The confidence figure
What the percentage beside a number actually means.
This deserves a full explanation rather than a footnote, because a charterer will test it against outcomes within a week — and an unexplainable score discredits every other figure in the product.
A description of the inputs
- How complete and how well-sourced the inputs were
- How much rests on verified data rather than a single circular
- How many values were assumed rather than known
- How many were adjusted by hand
- In short: how much of this figure does the system actually know
Not a forecast
- Not a probability that the voyage will earn this TCE
- Not a statement that the fixture is a good one
- Not a substitute for reading the workings
- A high score on a bad assumption is still a bad estimate
- Set a bunker price the market will not deliver and the arithmetic is confidently wrong
Which is why the figure sits beside a link to the workings rather than in place of them. The score tells you how much to check; the workings tell you what to check.
Failure
It will get things wrong. Here is what happens then.
Any vendor claiming otherwise is describing a demo. The useful question is not whether a system fails but whether it admits to it.
Nothing is dropped
A message that cannot be parsed goes to the Exceptions queue with the reason attached. It does not disappear, and it does not sit in a feed looking as though it was handled.
Nothing is guessed
A field the parser is unsure of is left empty and marked, rather than filled with a best attempt. An empty field prompts a person; a plausible wrong one does not.
Ambiguity stays separate
Where two circulars might or might not be the same cargo, they are not collapsed. A duplicate shown is a minor irritation; two different cargoes merged is a lost opportunity.
Accuracy is measured, not quoted
Extraction accuracy moves day to day with the traffic arriving. The current figure is visible in the product rather than printed here, which is the only place it can be honest.
Overrides
You can change any figure. It will note that you did.
Often you should. You may know the vessel, know the owner, or have been told something over the phone that has not reached any register. Judgement is the job, and a system that will not let you apply it is useless.
What it will not do is forget. The original is retained, the estimate is marked as user-adjusted, and anyone reading it later — including you, in six months, defending a claim — can see which figures were the system’s and which were yours.
Estimate marked user-adjusted. Warranted and observed figures retained. If the vessel underperforms, the evidence is already assembled.
Agents
Being built. They will stop at drafting.
Agents are in development and will be released progressively rather than announced all at once, coordinated by Captain AI — a master agent that orchestrates the workflows and manages the specialist agents beneath it.
Two things are worth saying now because they are design decisions rather than roadmap items. An agent will prepare a reply, an estimate, a set of instructions or a claim — and it will not send, offer, counter or commit you to anything. And anything an agent produces opens to its inputs, exactly as every other figure does.
Your data does not train them
Your circulars, recaps, fixtures and rates are never used to train, fine-tune or evaluate any model. Parsing runs against your data; it does not learn from it.
One client’s information is never visible to another, including in aggregate or anonymised form. There is no shared cargo pool and no pooled learning.
We will not aggregate your fixtures into an index, a benchmark or a report for anyone else. We are not in the market intelligence business and will not enter it using your book.
The agent writes. You decide. That does not change as the agents get better.
Book a demo
Send us the circular you think will break it.
Not the clean one. The badly formatted one from the broker who abbreviates everything, or the recap somebody photographed. Where it fails is more useful to both of us than where it succeeds.
We reply within one business day. Material you send is deleted after the session.