Bills of lading, customs declarations, invoices and packing lists at volume — where the pilot passes on the clean scans and the project is decided by the small fraction that are rotated, handwritten, or in a language nobody planned for.
All Solutions by Industry →
Two numbers decide whether a logistics AI project works, and neither of them is accuracy on a curated sample. The first is cost per document, fully loaded, against the cost of the clerk it is meant to relieve. The second is what happens to the documents the system cannot read — not as a fallback, but as the primary design surface, because the exception path is where the operational reality lives.
So the architecture puts a cheap triage step before any expensive model call. Legibility, orientation, page count and template match are scored first: rotated pages are rotated, unreadable pages are rejected with a specific reason rather than sent to a model that will produce something plausible about them. Documents are batched by template rather than by arrival, which lets one prompt carry template-specific instructions and a smaller model do the work.
Validation is cross-field rather than per-field, because individually plausible values that are jointly impossible — a total that is not the sum of its lines, a delivery date before dispatch, a weight that contradicts the packing list — are the highest-yield check available and cost nothing to run. And exception queues are typed. illegible_scan, unknown_template and arithmetic_mismatch go to different people with different fixes, and the distribution across those types tells you where to invest next. The whole design is set out on the document pipeline reference architecture.
Bills of lading, customs declarations, commercial invoices, packing lists and delivery notes read by multimodal models, with triage before inference, cross-field validation and typed exception routing. Batch economics rather than synchronous calls where the latency is not needed.
SEE THE REFERENCE ARCHITECTURE →The queue that decides whether the project is loved or resented. Each exception carries its type, the field that failed, the page it failed on and the fix it needs, so a clerk resolves it in seconds instead of reopening the whole document.
SEE CROSS-PLATFORM WORKFLOWS →Demand signals, lead times and supplier performance assembled into a form a planner can argue with, with the assumptions visible. The model surfaces and explains; the plan is committed by a person who owns the consequence.
SEE SUPPLY CHAIN PLANNING →Status, ETA and clearance questions answered from the TMS and port feeds directly, with the source and timestamp of every figure shown. Anything touching a claim, a detention charge or a customs hold routes to a person with the history attached.
SEE SUPPORT AGENTS →A document pipeline is one of the few AI workloads where the business case is genuinely arithmetic rather than argument, which is why we start there. The unit is one document processed end to end, and the honest version of that number includes six lines: triage, the extraction call itself, validation, retries on the calls that fail or time out, the review minutes spent on the exception queue, and the standing engineering time to keep templates, prompts and evaluations current as partners change their forms.
Two of those are usually missing from a vendor estimate. Retries bill and never reach a user. Exception review is a real cost that scales with the exception rate, which is why a pipeline that extracts a little less but fails in a typed, quickly resolvable way often beats one with a higher headline accuracy and an untyped failure bucket. The method for building that model, including the lines most estimates omit, is on the cost model page.
The comparison at the end is against the fully loaded cost of the same document today — the clerk, the rework, the demurrage caused by a late declaration, the management time. If the pipeline does not clear that number at your realistic volume, it is not a project, and finding that out in a fortnight of measurement is cheaper than finding it out in month nine.
These are engineering controls we build into the architecture, not certifications. QQuantum.ai holds no security certifications, and any audit your customers or customs authority require remains yours — see our trust centre.
The honest answer is that any figure quoted before seeing your documents is about someone else’s documents. Accuracy on freight paperwork depends almost entirely on the mix — a book of clean electronic PDFs from three forwarders behaves nothing like a book of scanned originals from forty. What we do instead of quoting a number is build an evaluation set from a representative month of your own traffic, including the awkward lanes, and report per-field accuracy against it before the build is scoped. That measurement is the deliverable, and it is what any later claim has to be checked against.
They go into a typed queue, which is the part of the design that matters most. A failure labelled only as failed is not actionable; a failure labelled illegible_scan on page three, or arithmetic_mismatch between the invoice total and its line items, routes to a specific person with a specific fix. The distribution across those types is also the roadmap: it tells you whether the next investment is a better scanner at one partner, a new template, or another language.
Yes, through validated interfaces with a schema check on every write, and with the write scoped to the fields the workflow actually needs. We are deliberately conservative here — an extraction system with broad write access to an operational system is a much larger blast radius than the same system emitting a record for a queue, and the second design is nearly always sufficient.
Not by default. The pipeline runs in your own cloud accounts against your own stores. Where a hosted model is used for extraction, we record the provider’s retention and training terms in the architecture document at the time of writing, and you should confirm them against the provider’s current terms before relying on them. Where commercially sensitive rate or customer data must not reach a third party at all, an open-weight model on your own infrastructure is the design, and the capability trade-off is stated before you commit to it.
That is not the shape we have seen work. What changes is where the team’s attention goes: away from re-keying fields that a machine can read, towards the exceptions, the disputes and the partner relationships that actually decide whether a lane is profitable. A pipeline sized on the assumption that the exception queue needs nobody is a pipeline that will be switched off in its second quarter.
Bring a representative month of real paperwork, including the bad scans. Thirty minutes on the mix is enough to say whether the arithmetic works.
30-minute technical call · no deck
What we built, what stage it is at, and a link to the live thing wherever there is one.
yourOS is an operating system with an AI at its heart. You talk to it like a person, out loud or typed. It makes your calls, plans your days and works on your projects with a crew of agents, on your own private server. Our own product, built from the first commit; the waitlist is open.
Website, backend and app assets for Sonora, a free AI sound-therapy app on iOS and Android published by Cognitive AI.
Brand, website, prototype, apps on five platforms, the AI backend and the algorithm-intelligence pipeline for Podswap, a social growth engine that runs without a human in the loop.
In closed beta: Roxy, a private AI assistant with permanent memory running on infrastructure the user owns. Brand, site, product, backend and conversational design.
Live at coherence.com: brand identity generated with AI agents we directed, the motion system and website for the group behind Sonora, AboutAI, WakeUp and MudShadows. Agent work remains in build.
Live at domainbooth.com for DomainBooth, a premium domain brokerage: a new animated mark, every page redesigned and rebuilt with designed mobile layouts, and a custom CMS for its sales record, featured names, press and enquiries.
Apps, platforms and campaigns for these teams, long before the AI work. Every one is written up on the work page — what we did, and where our part ended.
