Exim Routes AI Document Processor

Exim Routes, a logistics and trade platform, receives bookings, invoices, bills of lading, permits and compliance certificates from shippers, carriers and agents every day. Each one arrived in a different layout and had to be read and keyed into the system by hand — slow, repetitive, and one typo away from a delayed shipment.

• Carriers, forwarders and customs authorities each use their own document layout and field names.
• Inputs ranged from clean PDFs to scanned slips, Excel sheets and forwarded emails.
• A single misread field — a container or booking number — can hold up a shipment at customs.
• Rule-based OCR and fixed templates broke every time a carrier updated its format.

DevSecCops built a LangGraph multi-agent workflow on AWS Bedrock (Claude Sonnet 5 and Haiku 4.5) that classifies each incoming document, routes it to a specialist extraction agent, and grades the result before it ever reaches the database. Agents, schemas and prompts are defined entirely in YAML, so onboarding a new document type is a config change, not a code release.

• Automatic intake from a monitored mailbox or direct upload, deduplicated by Gmail message ID and SHA-256 file hash so retries and re-sent emails never create a duplicate record.
• Converts PDFs, scanned pages, Excel and CSV to text, falling back to Bedrock vision OCR when a scanned page has too little extractable text.
• A classifier reads the document and hands it to one of five specialist extractor agents — booking confirmations, bills of lading, commercial invoices, customs declarations and compliance certificates.
• An optional human-confirmation step lets an operator correct a misclassified document before extraction runs.
• A grading model scores every extraction for groundedness and completeness on a 0–1 scale, feeding back field-level corrections and retrying up to 3 times below a 0.85 threshold.

Every extracted value is returned with its source location and checked against the original document text — anything that isn't actually printed on the page is dropped rather than guessed. Final results, source files and every LLM call are stored with a full audit trail in PostgreSQL and traced in LangSmith, so any extracted field can be traced back to the document and model call that produced it.

• 88–98% of printed fields extracted correctly across 8 real client documents, hand-graded against the source.
• Just 1 fabricated value across all evaluation runs — the grounding check catches the rest before they reach the database.
• 5 document types live today, each one a YAML file rather than a code change, so new formats onboard fast.
• Idempotent intake and bounded retries mean crashes, restarts and re-sent emails never create duplicate or stuck records.
• Per-document LLM cost and token tracing gives operations a clear view of processing cost as volume scales.