Case studies/Logistics and freight
A Los Angeles customs broker filing entries against a 42,000-line product master, with classification drift measured across the prior fiscal year. We implemented a pipeline that proposes HTS codes against the master itself, not the last shipment, with a confidence threshold set per commodity family.
| Part | Family | Proposed HTS | Master HTS | Conf. | Status |
|---|---|---|---|---|---|
| A-44201 | apparel | 6110.20.2010 | 6110.20.2010 | 0.97 | classified |
| A-44218 | apparel | 6203.42.4011 | 6203.42.4011 | 0.96 | classified |
| S-30914 | industrial | 7308.30.5050 | 7308.30.5050 | 0.98 | classified |
| E-71220 | electronics | 8517.62.0090 | 8517.62.0090 | 0.95 | classified |
| A-44307 | apparel | 6109.10.0040 | (new SKU) | 0.93 | master pending |
| C-12508 | chemicals | 3824.99.9297 | 3824.99.9297 | 0.94 | classified |
| A-44119 | apparel | 6110.30.3053 | 6110.20.2010 | 0.91 | drift, filer review |
| P-58432 | parts | 8708.29.5060 | (new SKU) | 0.92 | master pending |
| S-30922 | industrial | 7228.30.8050 | 7228.30.8050 | 0.99 | classified |
At a glance
A broker with a product master that already existed. The job was to make the master the authority, not a reference document filers consulted after the fact.
The engagement
The stack
ISO 27001 · ISO 9001 · DPA and NDA signed at kickoff.
Before, the filer desk
The filers were good at their jobs. The fastest HS lookup was also the wrong one, and nobody had time to argue with the clock.
When a filer needed an HS code under time pressure, the fastest path was to pull the last invoice from the same importer. The product master sat in the entry system, slower to search, and not everyone on the desk had direct access.
Pre-build baseline: approximately 3% of lines flagged by CBP audit for HTS inconsistency across the prior fiscal year.
Commodities sitting across two HS chapters (textiles finished vs fabric, chemicals named vs mixtures, parts of vs complete) drifted over six months. Each individual filer classified consistently; across the desk, the drift compounded.
Pre-build baseline: review cycle post-audit consumed approximately 2 full-time filers per quarter on rework.
Apparel classifications carry more ambiguity than industrial chemicals. The legacy workflow did not distinguish. A filer's review time was the same across a clean steel shipment and a mixed-fabric garment packet.
Pre-build baseline: no per-family confidence threshold in the prior workflow.
What we built
The pipeline follows the same five stages we run on every logistics engagement. The details below are the ones we wrote against this broker's product master, not a generic template.
Commercial invoices and packing lists arrive via importer portals, broker email, and EDI. The product master is read on the nightly export. All three normalised into a single entry record.
Each invoice line tagged with a commodity family (textiles, chemicals, electronics, parts) from the master. LlamaParse primary, Google Document AI fallback. Tag confidence below 0.92 holds the line.
Line description, quantity, unit value, country of origin, importer part number. Part numbers matched to the 42,000-line product master; unmatched parts held for filer review.
HS code proposed from the master, not the last shipment. Confidence threshold set per commodity family (apparel 0.95, industrial 0.90). Below threshold, the line holds for filer review.
Cleared entries filed to ACE / ABI via EDI. Source documents attached to the broker file. Exceptions surface in a named filer queue with the family and the confidence in plain English.
After, the numbers the filer desk signs off
Same filers, same importers, same licence. The pipeline proposed HS against the master and surfaced only the lines below the family threshold. The review burden narrowed without the desk changing shape.
Filers still own every borderline classification. They still sign off every below-threshold line. What changed is where their time goes. On a steel shipment, the pipeline files. On a mixed-fabric apparel packet, the filer reviews the flagged lines and files in minutes.
From the licensed desk
The master was always the source; the process never matched. The pipeline just closed the gap, and now the drift does not compound.
Compliance leadCustoms broker, Los Angeles
Handover
The engagement ends at a clean handover. Compliance runs the pipeline; Hexaa stays on call for a fixed retention period, then steps back.
Related cases
Each links to a named client, a named document, and the system the clean data lands in. We publish only what the client signed off to publish.
28,000 customs packets a month reconciled inside a four-hour clearance window. LlamaParse on intake, CargoWise as the system of record.
→Logistics · 2023Freight forwarder · commercial invoice variancePacket-level variance flagging against the packing list before filing. 8,400 packets a month, US-EU lane.
→Construction · 2026Mid-market GC · AIA A201 against house positions180-page MSAs triaged in 8 minutes against a 34-clause house taxonomy. LlamaParse and LLM clause classification.
→Free 30-minute call
You'll leave with a clear next step.
The invoice line carries a description and an importer part number. The master carries the canonical HS for that part. The pipeline compares the proposed HS against the master-of-record and flags any drift before the entry is built.
