Form-extraction pipelines
Primary submission pipeline
Attachment to filed record in 90 seconds. Per-field confidence scoring with a reasoning trace.
Read the caseAll case studies
Case study · Catalog reconciliation
Meridian Operations · multi-line declarations with regulatory thresholdsCatalog reconciliation: 14× faster, 91% match
A reviewer who used to spend an afternoon cross-referencing packing lists against a product catalog by hand now spends four minutes approving a pre-matched summary — and the catalog fills itself in as new items arrive.
0%
Packing-list line items matched
0x
Faster than manual reconciliation
0.00
Threshold-validation pass rate
Built with
The problem
A 60-line packing list took 90 minutes — and the threshold checks happened at the end
Coordinators copied lines from packing lists into spreadsheets, looked up codes in catalog tabs, and typed values into a third column. The same product appeared under three different descriptions across three accounts. Threshold violations only surfaced when the reviewer checked the final form, by which point the rest of the packet was already done. Rework was the default state.
- Line-by-line in a spreadsheetCoordinators copied each line from the packing list into a spreadsheet, looked up the product in the catalog tab, then typed the catalog code into a third column. A 60-line list took 45 to 90 minutes — and errors only surfaced when the reviewer checked the final form.
- Threshold violations caught at the wrong stepRegulatory caps on volume and percentage limits were checked at the end of the process, not during data entry. By the time a violation flagged, the coordinator had already finished the form. Rework was the norm.
- Description drift across accountsThe same product appeared under different descriptions depending on which account submitted: "750ml Cabernet Sauvignon" from one, "Cab Sav 750" from another. Without a per-account alias table, both required manual catalog lookup every time.
- No visibility into unmatched itemsWhen a line didn't match anything in the catalog, it was left blank and flagged with a sticky note in the PDF. Catalog curators had no systematic view of what was accumulating across packets; the catalog grew only when someone remembered.
Unmatched items · this week
accumulatingHartwell Trading Co.Cab Sav 750ml ×120alias drift
Westridge Imports750ml Cabernet Sauvignon ×96alias drift
Cascade LogisticsPinot Gris 1L ×48no match
Northbay GroupCabernet Sauvignon 1.5L ×24 — over 1L capthreshold flag
Olmsted & SonsCabernet Sauvignon, 750mlalias drift
Riverhead Co.Pinot Noir 750ml — new SKUno match
Same products · five spellingsNo alias table
The pipeline
From inbox to verified record in one pass
Six steps. Catalog match and threshold check happen before the reviewer ever sees the packet.
01Trigger
Account submits packing list
PDF or scanned image uploaded via Zoho Desk ticket; attachment URL queued for processing.
Zoho Desk
02Detect
File type detected; scans OCRed
PDFs proceed to extraction; scanned images run through CloudConvert OCR to produce a searchable PDF first.
CloudConvert
03Extract
Line items pulled with quantity, unit, description
Claude reads the packing list and returns each line as a typed record: description, quantity, unit, value per unit, with source spans.
Claude
04Match
Catalog matched per account
Description fuzzy-matched against the account's catalog and alias table; exact matches auto-resolve; unmatched items go to the curator's enrichment queue.
05Validate
Threshold rules per line
Volume per unit, percentage limits, and restricted-product flags checked against the account-specific thresholds in the catalog. Violations held for reviewer action.
Azure Functions
06Confirm
Reviewer approves the summary
Reviewer sees one screen: matched items grouped, threshold flags shown inline, unmatched items separated. Approves matched rows in bulk; routes individual rows back if needed.
High confidence
Bulk approval in a four-minute median
91% of line items match the catalog at or above the confidence threshold; threshold checks pass at 0.94; reviewer approves the packet in a median of 4 minutes and the form is submitted. Matched items raise catalog confidence scores, improving the next packet from the same account.
Low confidence
Row-level routing into the enrichment queue
Unmatched items land in the catalog enrichment queue with the extracted description and source span. Threshold violations are held with the specific rule and the measured value displayed inline. Reviewers resolve one row at a time rather than re-doing the whole list.
Step 04 — Extraction
What Claude saw → what the system received
Scroll triggers the live extraction. Each region Claude reads on the document maps into the structured form on the right.
Source attachment
packing-list-MO-50412.pdfClaude
Target schema
Line items × 60✓ extracted
Cabernet Sauvignon—
✓ extracted
120—
✓ extracted
750ml bottles—
✓ extracted
$18.50—
✓ extracted
2204.21.5000—
0 / 5 fields populated✓ ready for review
Step 06 — Accuracy review
The system tells you when it's unsure
A second Claude pass scores every field against the source spans. High-confidence results post automatically. Anything ambiguous routes to a human with the reasoning attached.
Field-level confidence
Pass 2 — Claude self-review
HighMediumLow
Line descriptionCabernet Sauvignon
97%High
Quantity120
95%High
Unit of measure750ml bottles
92%High
Declared value per unit$18.50
89%High
Tax classification2204.21.5000
58%Low
Routed to human review. Three candidate codes match the description with overlapping coverage: 2204.21.5000, 2204.21.6000, and 2204.21.8000. The description alone is insufficient to distinguish between them; routing to reviewer to choose.
4 of 5 fields cleared the 0.85 threshold
model: ClaudeThe stack
Boring tech, glued together well
Each vendor handles what it's best at. Aisyst owns the orchestration layer in between.
Zoho Desk
Inbound submission queue; reviewer confirmation written back as ticket annotation
Claude
Extracts line items with source spans, then reviews per-field confidence
CloudConvert
OCR preprocessing for scanned packing lists before extraction begins
Azure Functions
Hosts the extraction pipeline, catalog matching, threshold validation, and enrichment queue writes
Postgres + Drizzle ORM
Per-account catalog with aliases, classification overrides, threshold rules, enrichment queue
Third-party logos are trademarks of their respective owners and appear here only to indicate integration.
Outcomes
What changed for the reviewer
Approve a pre-matched summary in four minutes instead of reconciling lines by hand for an afternoon. The catalog itself gets sharper every week as unmatched items accumulate in the enrichment queue.
0%
Line items matched to catalog
0x
Faster than manual reconciliation
0.00
Threshold-validation pass rate
0 min
Median per-packet reviewer time
Watch the per-account match rate, not the global one
A new account starts low — sometimes below 50% — because the catalog is empty. As reviewers resolve unmatched items and the curator adds entries, each subsequent packet from that account matches more lines automatically. If the rate flatlines for a new account after four weeks, the curator hasn't kept up with the enrichment queue. If the rate drops for an established account, the submitter has changed how it describes products — a description- drift problem the alias table fixes.