Didn’t Find The Answer You’re Looking For?
Call us, email us or reach out on social media.
Partnering with Aya Data to Power Desteia’s High-Precision Customs AI
Engagement Snapshot
Domain | Supply Chain & Global Trade Technology |
Core Initiative | Document intelligence and data annotation for Mexican customs compliance |
Engagement Duration | 3+ years (completed) |
Data Volume | 600+ PDF bundles, 30,000+ pages (estimated from SOW) |
Document Coverage | 8 core document types, mapped to 13 taxonomy categories and sub-types |
Quality SLA | ≥95% file-level accuracy per batch |
Results at a Glance
Over 3+ years, Aya Data annotated an estimated 30,000+ pages of customs and transport documents for Desteia. Desteia reports a 20% gain in extraction precision over the partnership.
Metric | Result |
|---|---|
Extraction precision | 20% increase across the partnership |
Volume annotated | 600+ PDF bundles, 30,000+ pages |
Batch quality | ≥95% file-level accuracy SLA, enforced on every batch |
Launch cadence | 2 new document-type projects per month |
Manual review baseline | 25–30 minutes per document before automation |
Engagement length | 3+ years |
Client Overview
Desteia builds AI that reads and validates the documents behind cross-border shipments into Mexico. Under rules such as the mandatory Manifestación de Valor Electrónica (MVE), one filing error can mean a held shipment, an administrative penalty, or a supply chain delay.
To properly train its systems, Desteia required high-precision, verified training datasets produced at scale. Across a three-year collaboration, Aya Data executed boundary detection, categorization, and field-level parsing for upwards of 600 multi-document PDF packages-representing more than 30,000 pages-while consistently upholding a strict ≥95% file-level accuracy benchmark.
The outcome: Desteia reports a 20% increase in extraction precision, and now launches about two new document-type projects per month.
The Challenge: Accuracy That Didn’t Scale
Before the partnership, Desteia relied on a mix of fully manual review and an early-stage model. Error rates were very low, but only because volumes were low. A manual review took 25–30 minutes per document, which capped how far Desteia could grow.
Scaling the model meant solving three data problems:
- Mixed document bundles. Shipments arrive as multi-page PDFs or multi-tab spreadsheets with several document types in one file, such as invoices, bills of lading, packing lists, and Pedimentos together.
- Dense regional schemas. Mexican Pedimentos, Cartas Porte, and COVE filings have nested structures, multi-line item arrays, and regional tax calculations.
- Zero-tolerance compliance. One mismatch between an invoice, a transport record, and a tariff code can trigger a border hold. Training data had to be expert-validated.
The Data: Scale and Document Mix
Aya Data processed more than 600 PDF bundles from Desteia, most running 50 pages or longer, for an estimated 30,000+ pages in total. These figures come from the engagement SOW and are lower bounds.
Each bundle combined several of the eight core document types that sit behind a shipment into Mexico. Every page had to be separated, classified, and, where relevant, extracted field by field.
Stage | Document types |
|---|---|
Commercial | Proforma Invoice, Purchase Order, Commercial Invoice |
Shipping & origin | Packing List, Certificate of Origin, Bill of Lading |
Mexican customs & transport | Mexican Pedimento, Carta Porte |
These eight types mapped to a taxonomy of 13 primary categories and sub-types, covering ocean, air, and ground logistics, regional customs declarations, and trade agreements.
How Aya Data Built the Training Data
Every bundle went through a two-phase pipeline that turned unstructured PDFs into schema-validated training data.
- Phase 1: Boundary detection and classification
- Annotators reviewed each bundle page by page and tagged logical transitions, such as an invoice ending and a bill of lading starting, or two separate bills of lading in one file.
- Each separated document got a unique identifier and a label from the 13-category trade taxonomy.
- Phase 2: Schema-constrained field extraction
- Values were normalized into strict JSON types: ISO dates, ISO country codes, container identifiers, and Incoterms.
- Nested structures, including tariff classifications, line-item tax breakdowns, and duty additions and deductions, were extracted into structured arrays.
- Entity profiles (consignors, consignees, brokers) and transport data (vessels, carriers, tracking numbers) were structured so Desteia’s system could validate across documents automatically.
Quality Assurance & SLA
Every batch had to reach ≥95% file-level accuracy on a random-sample audit before delivery. One misclassified page or one field error counted against the whole file.
Control | How it worked |
|---|---|
Batch audit | Random sampling per batch; ≥95% file-level accuracy required to pass |
Human-in-the-loop review | Specialists corrected boundary misalignments and field errors in pre-annotated model outputs |
Edge-case log | Unusual layouts, rare regional filings, and ambiguous trade terms were logged and fed back into the guidelines |
Desteia also credits Aya Data with consistently meeting delivery time frames, which left Desteia room to verify data before release (see Client Voice).
Guideline Evolution: A Two-Way Feedback Loop
Desteia owned the annotation guidelines; Aya Data’s production work kept them sharp. The loop ran in every phase of the engagement:
- Desteia supplied the baseline guidelines: regulatory schemas, business logic, and known edge cases.
- Aya Data’s annotators flagged ambiguous layouts and unresolved edge cases during production.
- Desteia updated the guidelines with added context and clarified rules.
- Aya Data applied the revised guidelines to the next batch.
This cycle cut ambiguity over time and kept output quality consistent as new document layouts appeared.
Business Impact
Desteia’s extraction precision rose about 20% over the partnership, and its launch cadence for new document types reached roughly two projects per month.
Area | Before Aya Data | After Aya Data |
|---|---|---|
Review approach | Fully manual review plus an early-stage model | Model trained on expert-validated data |
Manual review time | 25–30 minutes per document | Open: needs Desteia’s current figure |
Extraction precision | Baseline | 20% higher |
New document-type launches | Limited by manual capacity | 2 projects per month |
Error rate | Very low, at low volumes only | ≥95% file-level accuracy held at 600+ bundles |
Ongoing Value: Aya Data as Desteia’s Strategic Partner
Aya Data’s involvement goes beyond initial dataset creation, providing long-term strategic value to Desteia:
- Ongoing Model Optimization: Aya Data continues to annotate complex edge cases and newly emerging document layouts, preventing model drift and ensuring Desteia’s AI adapts to evolving supply chain formats.
- Regulatory Agility: As Mexican and international customs frameworks shift, Aya Data rapidly adapts to updated guidelines, enabling Desteia to ship compliant features faster.
- Accelerated Time-to-Value: By combining high-precision data annotation with rigorous human-in-the-loop validation, Aya Data directly reduced manual review cycles for Desteia’s end customers.
Client Voice (New Testimonials )
Contact Aya Data today to discover how we can automate your data extraction workflows.
Disclaimer: Aya Data respects client confidentiality and will not share client details or project information without the client's approval.
