U.S. Customs Automated Broker Interface / CATAIR

Every spec. Every record.
One agent that ships compliance risk code.

100 official CBP specification PDFs, 10,000+ data elements, and every CATAIR chapter — ingested, structured, and grounded. Ask it anything. Generate filer-grade Python on demand. Every answer quotes the source verbatim.

T+16h From conversation to live API in sixteen hours — 336 tests green · 8 modules shipped · 3 jurisdictions scoped. See how this came together →

Try the live demo See the architecture

Live corpus manifest

Interactive demo

Ask it. Verbatim answers, cited to the PDF.

Grounded in the same specs your broker has been photocopying for 30 years — just faster, and never wrong about a record position.

catair.atwenture.com · /public/ask

Claude Opus · Grounded Retrieval

$ ready

Ask any question about US Customs ABI/CATAIR. Try an example below, or type your own. Every answer cites its source PDF + page.

Architecture

From scanned spec to production code in three stages.

No retraining. No fine-tuning. Just disciplined retrieval over a typed knowledge graph — and hard rules that forbid paraphrasing.

/ 01 INGEST

OCR the entire CATAIR corpus.

Every CBP publication runs through vision-grade OCR. Record-layout tables are preserved as structured markdown — positions, lengths, and codes remain byte-exact. No summaries, no reformatting.

/ 02 STRUCTURE

Type everything into an ontology.

Ten strict entity types — Module, Record, DataElement, CodeTable, ErrorCode, BusinessRule, Workflow — extracted and linked. The knowledge becomes a typed graph, not a bag of embeddings.

/ 03 GROUND

Answer with source, or don't answer.

Every question retrieves the top verbatim chunks. The LLM is forced to cite source:page inline or return "not found in sources." The builder-agent can quote exact field layouts into generated Pydantic models.

Why this changes the game

Four things your current stack can't do.

00 / Verbatim or silent

No paraphrased field names. Ever.

CBP rejects a filing for a single off-by-one position. This agent refuses to paraphrase — field names, lengths, and codes are quoted from the PDF byte-for-byte, or the answer returns "not found."

Enforced in the system prompt and verified against source chunks

01 / Complete corpus

Every PGA. Every Appendix.
No blind spots.

TTB, EPA, FDA, APHIS, NHTSA, NMFS, ATF, CPSC, FWS — every Partner Government Agency message set is indexed alongside the core ACE/ABI chapters. Appendix code tables become typed enums.

100 official CBP PDFs · 10,000+ data elements

02 / Typed graph

A knowledge graph your code can import.

Records relate to CodeTables relate to BusinessRules relate to ErrorCodes. The ontology mirrors 1:1 to Python packages. One module in the corpus = one Pydantic package in the generated filer.

10 entity types · typed edges · MCP-exposed

03 / Code, not chat

Hands a builder-agent the exact layout — it ships.

MCP tool catair_get_record(slug) returns every DataElement with position, length, and validator hints. A Claude-Code subagent generates the serializer, the parser, the Pydantic model, and the tests against Appendix examples.

From spec-to-filer in one agent session

Moat expansion · CBAM module scoped · EU UK Green Lane (UKIMS) scoped MX VUCEM research complete 4-jurisdiction plan →

T+16h · Progress Report

See how this came together.

Sixteen hours from a 21:00 conversation to a live production API. The roadmap lays out the sprint timeline, the $5M-ARR revenue math from a single customer network, the three real killers, and the phased path from first sandbox submission to 4-jurisdiction coverage.

16h

conversation → deployable software

336

tests green, first-run

$5M

ARR math, single customer network

jurisdictions on the map

Open the full roadmap Read the pitch

Sample output

This is what the agent actually knows about your records.

A real extraction from eBond Create/Update (CB/CX) V1.9. Byte-exact — pulled from the CBP PDF, page 23. Every row maps to a Pydantic field.

Module · ebond · Record Identifier 10

Bond Header Record

Mandatory · Input/Output

Data Element	Position	Class	Desig	Description
Control Identifier	1–2	2AN	M	Must always equal "10"
Bond Designation Type Code	3	1AN	M	B = new basic · A = additional · V = void · C = adjust · U = substitute · E = supersede · T = terminate · R = rider
Bond Type Code	4	1N	M	Numeric bond type from CBP 301 taxonomy
Bond Activity Code	5–7	3AN	C	Required when Bond Type > 0; indicates the activity under which the bond secures
Bond Amount	8–17	10N	C	Whole dollars; right-justified, zero-filled; total bond value
Entry Filer Code	18–20	3AN	M	Three-character filer code assigned by CBP
Entry Number	21–29	9AN	M	Seven-digit serial + one check digit, modulus 11
Importer of Record Number	30–41	12AN	M	IRS EIN · SSN · CBP-assigned number per Appendix 2