Solution · Logistics
Convilyn Pre-Clearance Packet Agent
Pre-clearance preparation and review-packet generation. Sources integrated, discrepancies clarified, missing items exposed — the final call stays with a licensed broker.
Requirement confirmed
Fast delivery
Continuous security
How adoption works
Three phases, each ending in something you can sign off
An adoption project fails on scope, not on models. Each phase below closes with a written artefact, so the next one starts from an agreement rather than an assumption.
Requirement confirmed
We inventory the documents and systems already in use, agree which fields matter and how each is read, and draw the line where a person reviews rather than the workflow decides.
Output of this phaseA written scope note naming every field, every reading rule, and every human checkpoint.
Fast delivery
The workflow is assembled from generic platform tools rather than built for one customer. Real sample documents run first, and volume follows once the sample set is clean.
Output of this phaseA running workflow on real sample documents, with the cost per run measured before scale-up.
Continuous security
Every run leaves a tool-invocation record. Retention windows, deletion behaviour and the sub-processor list are published pages that a security reviewer can check without asking us.
Output of this phaseAn audit trail per run, and a documented retention, deletion and sub-processor position.
Scope of the exchange
What you hand over, and what comes back
A pre-clearance engagement is scoped by document type and by who signs. Both sides of that are named here before any work starts.
- Your organisation providesEmail threads, scanned faxes and supplier-portal downloads for the shipment, in whatever versions they arrived.Convilyn deliversFiles grouped by shipment, with threads that span several shipments flagged rather than silently split.
- Your organisation providesInvoices, packing lists, bills of lading and certificates, including revisions that superseded an earlier copy.Convilyn deliversEach document classified and version-tracked, labelled with its source page. Superseded and revised copies are shown together, never silently picked.
- Your organisation providesThe fields your filing system needs, and the checks a broker would otherwise run by hand.Convilyn deliversEvery field extracted with a source-page reference, and cross-document contradictions surfaced as a comparison rather than resolved by a guess.
- Your organisation providesThe packet layout your filing system or your customer expects.Convilyn deliversPDF, Excel or JSON assembled in that layout and ready for the filing system.
- Your organisation providesThe licensed broker who signs off, and the point in the process where they do.Convilyn deliversA tool-invocation record for every run. The final call stays with the broker, and each confirmation or override is logged.
From scattered email to review-ready packet
Same shipment, two routes. The manual path stitches across mailboxes, inboxes, and spreadsheets; the AI path keeps every field source-anchored and routes the decisions back to the broker.
Manual workflow06 steps
Receive client documents
Email threads, scanned faxes, supplier portals — files arrive in many places, in many versions.
Triage by type, label by hand
Track versions in a spreadsheet(this step no longer happens)
Old vs revised invoice. Email instruction vs file timestamp. Easy to lose track.
Read each document, transcribe fields
Cross-check for contradictions
Invoice total vs line-item sum. Packing weight vs B/L weight. Catch inconsistencies before customs does.
Compile packet, hope nothing missed
Final review depends on the operator noticing what is missing — there is no automated cross-check.
Convilyn workflow05 steps
Upload documents
Drag in the full thread. The agent groups files by shipment and flags multi-shipment threads.
AI classifies and version-tracks
Invoices, packing lists, B/Ls, certificates — labelled with the source page; old vs revised flagged, never silently picked.
AI extracts and flags inconsistencies
Every value carries a source page reference. Cross-document contradictions surface as a diff, not a guess.
Reviewer scans the diff and conflicts
The licensed broker confirms or overrides. Every decision is logged.
Approved packet exports
PDF / Excel / JSON, ready for the filing system.
Expected result
One shipment, end to end
The same shipment, from the state it actually arrives in to the packet a broker reviews.
Worked example
A shipment arrives as 14 attachments across three email threads: two invoice revisions with no version marker, a packing list, a bill of lading, and a scanned certificate of origin naming a different consignee.
Input
- 3 email threads, 14 attachments, no consistent file naming
- 2 invoice revisions, neither marked as the current one
- 1 certificate of origin, scanned at low resolution
Deliverable
- A review packet in your filing layout, every field carrying its source page
- A conflict list: invoice total against line-item sum, and consignee name across two documents
- A missing-document list, checked against the filing requirement rather than against memory
What to measure
- Time from inbox to review-ready packet
- Fields re-keyed by hand per shipment
- Conflicts found before filing rather than by customs
- Packets returned for a missing or superseded document
System integration
Build once, call the same workflow from the web, the API or an SDK
A workflow is built once in Builder. All three call surfaces point at that one workflow, not three copies to keep in step.
In
- Direct upload
- Google Drive
- Dropbox
- URL import
- Dedicated inbox
Workflow substrate
- Excel
- Word
- JSON
Out
Call surfaces
Web interface, REST · Python SDK · TypeScript SDK. Agent workflows stream events over WebSocket; deterministic conversion is polled.
Where it runs
Workflows run in the cloud. Ainalyn for desktop (beta) handles image, PDF and OCR conversion on the machine; convilyn.local converts offline with no key and no network.
Your own tools
You can host a tool server and have the platform call into it. Every outbound request carries a HMAC-SHA256 signature and a timestamp, so you can verify it came from us.
URL import accepts https only, does not follow redirects, refuses addresses that resolve to internal ranges, and has a size limit. Download links are short-lived presigned URLs and can be reissued.
Why Convilyn
The substrate is ours, and the guarantees are checkable
Document automation projects stall at security review. Every item in this section can be checked before you talk to us.
Decision boundary
Eight gates decide in code, against fixed thresholds, in a fixed order. A model cannot skip one, and cannot argue a threshold up.
- RedactionSensitive content is masked before it reaches a model.
- Budget ceilingA per-request spend cap. Work stops at the limit.
- Retry policyFixed backoff, so a rate limit never becomes a stampede.
- Phase legalityA tool runs only in the phase the workflow allows.
- Cycle detectionThe same call repeated in a row aborts the run.
- Output reject limitThree rejected outputs and the run stops retrying.
- Tool permissionWriting to an outside system needs granted authority.
- Code rescue admissionWhether a code rescue may run is configuration, not a judgement.
Extraction is held to the same kind of rule: a field the source does not contain is left empty. A grounded blank is the correct answer; an invented value is not.
Privacy is the default, not a setting to find
Model providers do not train on your content. Runs started with an API key are excluded from Convilyn's own model work in code, and fail closed — a refusal by default, not a policy statement.
Integration flexibility, because the constraint is usually on your side
The interfaces are listed above. The other half of that commitment is this: if a requirement is beyond us, we say so during the assessment rather than at go-live.
Model training
Inputs and outputs sent through AWS Bedrock are not used to train Anthropic's models (see Bedrock's data protection terms). Runs started with an API key are excluded from Convilyn's training data, enforced in code.
Deletion
Files can be deleted immediately through the API, by the uploader only. A file attached to a running workflow is protected until that run reaches a terminal state.
Encryption and isolation
TLS in transit, encrypted at rest. Storage paths are isolated per account and every read and delete checks ownership — guessing an identifier does not reach another account's file.
Redaction
Rules match on field name and value shape, so content inside fields like resume_text is caught too. It does not rely on a model to judge what is sensitive.
Sub-processors
Every sub-processor handling customer content is listed publicly, with its role and processing region. That page is the current list at all times, and the address for questions about it is on the page.
Retention
- 1 hourSource files and download links
- 7 daysWorkflow run state
- 30 daysDeliverables
The cleanup pass runs every 15 minutes, which is why the shortest horizon is about an hour rather than exactly one. A file still attached to a running workflow is kept until that run finishes. Any file can be deleted immediately through the API.
Processing regions are fixed today: file storage and workflow execution in Tokyo (ap-northeast-1), model inference in us-east-1, and OCR in Singapore (ap-southeast-1). The region cannot be selected per request. Raise a data-residency requirement during the assessment and we will say plainly what is and is not possible.
Common questions
Questions that come up before a rollout
How long does a rollout take?
The requirement phase is usually the longest part, and it scales with how many document types and exceptions are in scope. Delivery runs against real shipment documents rather than a pilot environment, so a working output arrives before the full document set is covered.
Do we have to change our filing system?
No. Documents arrive by upload, cloud-storage import, URL or a dedicated intake email address, and the packet comes back in the layout you already file with, plus PDF and JSON. If you prefer a direct connection, the REST API and the Python and TypeScript SDKs are available.
Will customer shipping documents be used to train a model?
No. Invoices, bills of lading and customer data never become training data — not for the model provider, and not for us.
What happens when a value is read wrong?
Conflicting values are shown as a comparison instead of being resolved silently, and a field with no support in the source is left empty rather than filled. The licensed broker confirms or overrides, and every decision is logged.
How do we start?
The documents for a few real shipments, plus the packet layout you file with today.
Start with your own shipments
The assessment comes back as a written scope statement naming the fields, the reading rules and the broker's review points. No charge.
Book an adoption assessment