02 — Practice

Custom AI for the paperwork freight runs on.

Forwarders, customs brokers, importers and distributors — anywhere from one person to twenty, re-keying the same shipment out of a PDF and into a system three times over. We build the thing that reads the document, checks it against the other, and puts the result where you already look. Fixed price from $99, agreed before we start.

What it costs

Start at $99.

Nobody should hand four thousand dollars to a company they found on the internet last week. So the first thing we build for you is small and cheap enough to be a test rather than a decision. Every price is on this page, before any call.

$99 One small tool

Three days

For a business of one or two people with a single job that keeps eating an afternoon. A script that files documents the way you would have. A spreadsheet that fills itself from the PDFs sitting in a folder. One input, one output, working by the end of the week.

Ordered from the form — no call needed, and none scheduled
One round of changes after you have tried it
It does not connect to your other systems at this price
Sometimes there is AI inside it, sometimes it is just software

Past that, three sizes of project. Each one is a fixed number agreed before any work begins, and it does not move unless you ask for something we did not agree on. No hourly billing, no monthly fee, nothing to cancel afterwards.

$1,999 One task: One document type in, one place for the result to land. Commercial invoices into your spreadsheet, arrival notices into a dated list, bills of lading into the fields your TMS expects.
$3,999 A workflow: Two or three document types that have to agree with each other — invoice against packing list, discrepancies reported rather than summarized — plus a simple screen your operations team actually opens.
$5,999 Something harder: Documents arriving straight out of an inbox, formats from a dozen carriers, or reconciliation across a month of shipments. The jobs where the hard part is volume and variation rather than any single document.

Every project price includes the measurement — before we hand over, we run the system against a sample of your real files and tell you what changed. If the honest answer turns out to be that an off-the-shelf product does this for $30 a month, we will name the product instead of quoting you.

The documents

Four documents, one problem.

Anyone can demonstrate a model reading an invoice. The gap between that demonstration and something an operations team stops double-checking is specific to each document, and that gap is where the work is. A forwarder, a customs broker and an importer all handle these four — so what we build for one of you usually fits the next.

Commercial invoice

The work

The document that decides duty. Values, HS codes, Incoterms, parties, currency — arriving as a clean PDF export, a crooked scan, or a photo somebody took on a phone at a warehouse.

Where generic AI falls over

Every shipper formats it differently, and a general-purpose model reads it correctly most of the time. Most of the time is exactly the problem: a transposed value or a misread HS code does not announce itself, it surfaces weeks later as a customs query. What is needed is extraction that scores its own confidence field by field and flags what it is unsure of, rather than filling the gap with something plausible.

What a build usually takes

Layout-aware OCR · field-level confidence scoring · HS code validation against a reference table · deterministic arithmetic checks

Packing list

The work

Cartons, weights, dimensions, marks and numbers — checked line by line against the invoice before anything moves.

Where generic AI falls over

The value here is not in reading the document, it is in the comparison. Line counts that do not match, weights that do not add up, a carton on one page and not the other. Ask a language model to “compare these two documents” and it will hand you a fluent summary that quietly misses the discrepancy. The comparison has to run as code over extracted fields, and the output has to be a list of differences rather than a paragraph.

What a build usually takes

Table extraction · line-item matching · deterministic reconciliation · a discrepancy report, not a summary

Bill of lading and air waybill

The work

Shipper, consignee, notify party, container and seal numbers, port pairs, dates — the fields that get re-keyed into a TMS one shipment at a time.

Where generic AI falls over

Layout varies by carrier, so a single template breaks by the second client. But container numbers carry a check digit, which means a misread is detectable rather than silent — the system can know it got one wrong. Building that check in is the whole difference between something you can leave running and something a person has to proofread anyway.

What a build usually takes

Carrier-specific templates · ISO 6346 check-digit validation · entity matching against your existing contacts

Arrival notice and rate confirmation

The work

Charges, free time, the last free day, appointment windows — the numbers that decide whether you pay demurrage or not.

Where generic AI falls over

These land as email attachments at all hours, and the cost of missing one is measured in days of storage. There is nothing clever about getting them out of an inbox and into a dated list, which is precisely why nobody has built it for a four-person office. It is worth more than most of what gets sold as AI in this industry.

What a build usually takes

Inbox intake · charge extraction · date normalization · a dated list your team already looks at

Who this is for

Four businesses, the same paperwork.

None of you would describe yourselves as the same industry. You all spend part of every day moving the same fields out of a PDF and into a system.

Freight forwarders: From a one-person shop to a team of twenty, with a shared inbox and a TMS or spreadsheet that everything gets typed into twice.
Customs brokers: Entries built from documents that arrive in whatever shape the shipper felt like sending.
Importers and exporters: Running your own shipments and reconciling what arrived against what you were invoiced for.
Wholesale distributors: Purchase orders and supplier documents in the same formats, with the same re-keying underneath.

The toolkit

A language model is one instrument, not the orchestra.

Since 2023 almost every proposal in this space has been a language model with a prompt attached. Plenty of these problems are solved better, cheaper and far more reliably by techniques that predate the current wave — and knowing when to reach for which is the engineering.

Computer vision: Defect detection, measurement, reading a gauge or a label. Orders of magnitude cheaper and steadier than asking a multimodal model to look at the same frame.
Speech: Ambient transcription where the alternative is a professional typing up notes after the fact.
Document AI: Layout analysis, table extraction, OCR across forms nobody designed to be machine-readable.
Classical machine learning: Anomaly detection, forecasting, risk scoring. Decades old, well understood, and frequently the correct answer.
Retrieval: The part that decides whether a language model sees the right passage — and where most failures blamed on the model actually originate.
Language models: Drafting, summarizing, classifying, and calling the tools above. One instrument in the set, not the set.

How we engage

Four things that follow from the price.

Built on something we have built before

At these prices nobody can afford a system invented from nothing, and you should not want one. We start from parts we have already shipped — extraction, comparison, validation — and fit them to your documents and your steps. Less to go wrong, and it arrives in days or weeks rather than quarters.

One price, agreed before we start

You see the range on this page and a fixed number before any work begins. No hourly meter, no change-order theatre, and no discovery call whose real purpose is finding out what you can afford.

It runs without us

We hand it over and step back — no retainer, no monthly fee, nothing to cancel later. That constraint shapes the build: few moving parts, plain-language notes, and nothing that quietly needs us in six months. If you do want us again, it is another fixed price for a defined piece of work.

We will tell you not to buy it

A fair share of what people ask us for is already solved by a $30-a-month tool, and we would rather name that tool than take the project. It costs us the occasional invoice and saves you a month and a great deal more than we would have charged.

Build decisions

What separates a demo from a deployment.

The same five choices recur in every engagement, whichever field it is. Getting them wrong is why so much industry AI stalls after the pilot.

Confidence before convenience: Every extracted field carries a confidence score, and anything below the line is flagged for a person rather than filled in with something plausible.
Arithmetic belongs in code: Totals, weights, carton counts and duty are computed by deterministic code, never by a model. It decides what to check; it never does the sum.
Flagging beats guessing: A field marked “unsure, look at this” is worth more than one filled in correctly nine times out of ten, because the tenth time nobody catches it.
Inside the tools already in use: Output lands in your spreadsheet, your shared drive, the TMS you already pay for. Anything that needs its own login gets abandoned by the second month.
Measured on your own documents: Before we hand over, we run it against a sample of your real shipments and give you the numbers: how long the task took before, how long now, and how often someone has to correct it. Not a benchmark score — yours.

A system that says I am not sure about this one is worth more than a system that is right slightly more often.

What we optimize for

Re-keying the same shipment three times?

Send us a couple of paragraphs and, if you can, two or three real documents with the sensitive parts blacked out. You will get an honest read on whether it is worth building, which of the three prices it falls under, and how long it would take.

Start a conversation How we work