Chunghwa Post Customer story
Put a US code and duty on every parcel.
When US de minimis closed, every parcel bound for the US needed a real HTS code and import duty. A model on Chunghwa Post’s own server now answers in under 400ms.
- Parcels classified per dayManual250K+Peak capacity on their server
- Latency per parcel (p50)Minutes<400msBatch call, on-premise
- Auto-accepted at QA96%+The desk reviews exceptions only
- Per callUS HTS + dutyClassification, code and duty together
Runs on Chunghwa Post’s own server. No parcel data leaves their infrastructure.
Results after cutover. Parcels classified per day: Manual to 250K+ (Peak capacity on their server); Latency per parcel (p50): Minutes to <400ms (Batch call, on-premise); Auto-accepted at QA: 96%+ (The desk reviews exceptions only); Per call: US HTS + duty (Classification, code and duty together).
What happened at Chunghwa Post?
A three-month build put a classification-and-duty model on their own server, then a two-month rollout moved live parcels onto it.
Chunghwa Post is Taiwan’s national postal operator, and a large share of its parcels go to the United States. When the US de minimis exemption closed, each of those parcels needed a real US HTS code and import duty, and the manual classification desk became the bottleneck. The parcel data also had to stay on Chunghwa Post’s own infrastructure. GingerControl built the server and deployed a classification-and-duty model on it in three months, then rolled it out over two. It is sized for 250,000+ parcels a day at under 400ms each, with 96%+ auto-accepted at desk QA.
- Engagement
- On-premise model + server build
- Industry
- National postal operator, Taiwan
- Lane
- Taiwan → US parcels
- Build window
- 3 months
- Timeline
- 5 months: 3 to build, 2 to roll out
Case file for Chunghwa Post: Engagement On-premise model + server build; Industry National postal operator, Taiwan; Lane Taiwan → US parcels; Build window 3 months; Timeline 5 months: 3 to build, 2 to roll out.
Why did the desk become the bottleneck?
Parcels leave in waves of tens of thousands, and no manual desk can code and cost them that fast.
- When US de minimis closed, every US-bound parcel needed a real HTS code and import duty before it could enter.
- Parcels leave in waves with a thin description and little else. Latency, not accuracy, set the pace.
- More people raised cost without fixing speed, and national parcel data could not go to an outside cloud.
“Every parcel we send to the US needs a real code and a real duty figure before it ships. We needed that automatically, at our volume, on our own infrastructure.”
US de minimis closed: every parcel needs a US HTS code and import duty
- Fast enough for the waveClear one wave before the next leaves. A careful officer can’t.Required
- Code and duty togetherA code with no duty figure just moves the work downstream.Required
- Data stays homeCitizens’ parcel data can’t be sent to an outside cloud.Required
Requirements list for US-bound parcels after de minimis closed: clear each wave before the next leaves, return the code and the duty together, and keep parcel data off any outside cloud.
What runs on their own server?
One call in; the product classification, US HTS code and landed import duty out, all inside Chunghwa Post’s walls.
- Two required fields per parcel, description and country of origin, so the build didn’t wait on cleaning every upstream feed.
- Batch calls sized for a full peak wave, returning results in manifest order.
- We provisioned the server and deployed the model on it. No parcel data leaves the building.
- Shadow mode against the desk on the hardest categories, then a staged cutover.
“We did not need a faster desk. We needed the full answer, code and duty together, running where the data already was.”
- Description
- Country of origin
- Batch, manifest order
- Sized for peak waves
- Product classification
- US HTS code
- Landed import duty
Into the clearance flow
Deployment diagram: inside Chunghwa Post’s infrastructure, parcel records with a description and country of origin go into the classification and duty model, which returns the product classification, US HTS code and landed import duty for each parcel into the clearance flow. No outside cloud is used.
What did it cost?
A fixed fee of NT$9M+, paid against five milestones, and no per-parcel fee after go-live.
- Pricing model
- Fixed scope, milestones
- Paid per accepted milestone, not by the hour
- Investment
- NT$9M+
- One-time: on-premise model + server build (≈ US$290K+)
- Ongoing cost
- No per-parcel fee
- On-premise, no cloud usage bill
Engagement budget: Pricing model Fixed scope, milestones, Paid per accepted milestone, not by the hour; Investment NT$9M+, One-time: on-premise model + server build (≈ US$290K+); Ongoing cost No per-parcel fee, On-premise, no cloud usage bill.
Share of the engagement effort, from standing up the server to the staged cutover.
On-premise server and model work: 65% of the effort
Bar chart of the engagement effort: server build and on-premise deployment 35%, model deployment and accuracy tuning 30%, integration into the clearance flow 20%, shadow mode and staged cutover 15%. The server and model work together is 65% of the effort.
- The classification and tariff engine already existed, so the work was deploying and tuning it on their hardware and building the server around it.
- Ongoing cost is optional support and updates, not usage.
How long did it take?
Five months: three to build, two to roll out, with a test the client owned at every milestone.
- Month 1Discovery and data contractMapped the parcel feed, agreed the two required fields, scoped the server.Deliverable Signed scope and input contract, fixed price confirmed.
- Month 2MVP: one call, code plus dutyA US code and import duty for a single parcel on real sample data.Deliverable Working single-parcel call on Chunghwa Post samples.
- Month 3Server build and deploymentProvisioned the server, deployed the model, scaled to batch for peak waves.Deliverable Model live on their server, clearing a full wave.
- Month 4Shadow mode and testingRan alongside the desk on live parcels without touching clearance.Deliverable Accuracy report against the desk, signed off to go live.
- Month 5Training, cutover, evaluationTrained the desk, moved live traffic in stages, checked against the baseline.Deliverable Cutover complete, desk trained, results signed off.
Milestones: Month 1, Discovery and data contract; Month 2, MVP: one call, code plus duty; Month 3, Server build and deployment; Month 4, Shadow mode and testing; Month 5, Training, cutover, evaluation.
- Product classification, US code and import duty in one call
- The model running on Chunghwa Post’s own server
- Batch endpoint sized for a full peak wave
- Reconciliation summary against the source manifest
- Parcels to destinations beyond the US
- Self-serve dashboard and analytics for the desk
- Inbound classification for parcels arriving in Taiwan
MVP scope. In the MVP: Product classification, US code and import duty in one call; The model running on Chunghwa Post’s own server; Batch endpoint sized for a full peak wave; Reconciliation summary against the source manifest. Phase two: Parcels to destinations beyond the US; Self-serve dashboard and analytics for the desk; Inbound classification for parcels arriving in Taiwan.
- Payment followed acceptance, not the calendar.
- No live parcel went through the model until shadow mode matched the desk on the hardest categories.
- Live traffic moved over in tranches, and the desk went from classifying every parcel to reviewing exceptions.
What changed after cutover?
The desk stopped being the bottleneck: waves clear at batch speed and people review only the exceptions.
US-bound parcels cleared through classification, two weeks before to eight weeks after cutover.
Line chart of parcels classified per day from 2 weeks before to 8 weeks after the on-premise cutover: about 7,000 to 8,000 a day by the manual desk, climbing to about 250,000 a day by week six.
Before the on-premise model, and after cutover.
- Classifying every parcel
- Duty lookups by hand
- Reconciliation and chasing
- Judgment and exceptions
Bar chart of the customs desk’s time. Before: classifying every parcel 60%, duty lookups by hand 20%, reconciliation and chasing 12%, judgment and exceptions 8%. After cutover: 8%, 4%, 8% and 80%.
- The ceiling moved from what the desk could handle to what the server was sized for, and it was sized for the peaks.
- Every parcel carries its US import duty from the same call that classified it.
- The desk’s expertise moved from volume to judgment on the genuinely hard parcels.
“Every parcel we send to the US now carries a real code and a real duty figure before it ships, from one call we can audit end to end, running on our own servers.”
| Metric | Before | After | Change |
|---|---|---|---|
| Classification latency per parcel | Minutes, manual desk | <400ms, batch | Bottleneck removed |
| Parcels classified per day | Limited by the desk | 250K+ capacity | Server sized for peaks |
| Where parcel data lives | Would leave for lookup | Stays on-premise | Data sovereignty |
| Accepted at desk QA | Manual baseline | 96%+ auto-accepted | Desk on exceptions |
| Output per parcel | Code only, duty separate | Class + code + duty | One call |
Before and after: Classification latency per parcel, Minutes, manual desk before, <400ms, batch after (Bottleneck removed); Parcels classified per day, Limited by the desk before, 250K+ capacity after (Server sized for peaks); Where parcel data lives, Would leave for lookup before, Stays on-premise after (Data sovereignty); Accepted at desk QA, Manual baseline before, 96%+ auto-accepted after (Desk on exceptions); Output per parcel, Code only, duty separate before, Class + code + duty after (One call).
Why was this more than a faster desk?
Three calls: code and duty as one answer, the model brought to their data, and capacity sized for the wave.
- Code and duty are one answerA US code with no duty figure just moves the bottleneck downstream.Decided
- The model went to the dataA national post can’t send citizens’ parcel data to an outside cloud, so the server was built inside their walls. No per-parcel fee either.Decided
- Sized for the wave, not the averageParcels leave in waves, so the model runs in batch and answers in milliseconds.Decided
Decision log: Code and duty are one answer; The model went to the data; Sized for the wave, not the average.