Every capability the platform delivers today, grouped by the job it does.
Every capability the platform delivers today, grouped by the job it does. Status legend: ✅ live · 🟡 partial · ⬜ planned.
Document intake
| Feature | Status | Notes |
|---|
| Upload invoice / packing list / BoL (PDF or image) | ✅ | Drag-drop into the Document inbox |
| AI invoice extraction (OCR → structured fields) | ✅ | Pulls supplier, buyer, line items, values, references via Claude |
| Email-packet intake | 🟡 | UI affordance exists; live email ingestion not wired |
| Auto-create shipment from extracted invoice | ✅ | Extracted data becomes a draft shipment |
HTS classification (the AI core)
| Feature | Status | Notes |
|---|
| AI classifies each line to a 10-digit HTS code | ✅ | Claude Sonnet, grounded on real HTSUS + CBP rulings |
| Calibrated confidence per line (0-100%) | ✅ | ≥92% auto-approves; below routes to broker review |
| Citable rationale | ✅ | Cites the CBP ruling it relied on ("per NY N014838") |
| Alternative codes offered | ✅ | Up to 3 alternatives with reasons |
| Compliance flags | ✅ | Section 301, PGA triggers, dominant-fiber rules, etc. |
| Broker override with reason | ✅ | Every correction recorded for the feedback loop |
| Learns from corrections | ✅ | Recent broker corrections become few-shot prompt examples |
Denied-party screening
| Feature | Status | Notes |
|---|
| Screen importer / consignee / supplier | ✅ | Runs automatically during classification |
| 8 US sanctions lists | ✅ | OFAC SDN, BIS Entity/Denied/Unverified, State AECA/Nonprolif, BIS MEU, DHS UFLPA |
| Fuzzy name matching (incl. aliases) | ✅ | pg_trgm strict-word-similarity; catches alias hits like "Wagner Group" → "PMC Wagner" |
| Risk bands (block / review / watch) | ✅ | Block-band hits route the shipment to hold |
| False-positive dismissal with reason | ✅ | Feeds screening calibration |
| 31,400 entities loaded | ✅ | Refreshed from OpenSanctions (daily upstream) |
Duty + landed cost
| Feature | Status | Notes |
|---|
| Base HTSUS duty rate | ✅ | From the ingested tariff schedule, with 8-digit fallback |
| Section 301 (Lists 1-4A) | 🟡 | Broad-stroke coverage (~90% of common imports) |
| Section 232 Steel + Aluminum | 🟡 | Stacks correctly with Section 301 |
| IEEPA Russia | 🟡 | Broad-stroke coverage |
| Section 201 Solar | 🟡 | Broad-stroke coverage |
| MPF + HMF fees | ✅ | Per 19 CFR 24.23 / 24.24, with caps |
| Landed cost rollup per shipment | ✅ | Shown on dashboard + drawer |
| UFLPA presumption flagging | ✅ | Flags rather than charges (it's a detention process) |
| Penny-accurate code coverage | ⬜ | Needs U.S. Note 20(b) parsing — scoped, deferred |
| Live tariff-action feed | ✅ | Federal Register ingestion — FR alerts refresh daily; HTS/CROSS/screening weekly |
Tariff intelligence
| Feature | Status | Notes |
|---|
| Live alerts feed | ✅ | Auto-refreshing program alerts; surfaces which shipments are affected |
| Filter by list type (HTS / UFLPA / PGA) | ✅ | |
| Severity filtering + snooze | ✅ | |
| HTS reference rate table | ✅ | Base + Section 301 by code |
| Auto-detection of new actions | ✅ | Federal Register poller — runs on a scheduled refresh (FR daily; HTS/CROSS/screening weekly) |
Workflow + state machine
| Feature | Status | Notes |
|---|
| Entry lifecycle: draft → ai_review → ready → filed → released → cleared | ✅ | Plus hold / exam / rejected branches |
| Auto-routing on classification | ✅ | High confidence → ai_review; screening block → hold |
| Broker advance / approve / hold actions | ✅ | |
| Activity log (full audit trail) | ✅ | Every AI, screening, broker, and system action |
AI evaluation (the feedback loop surface)
The evaluation surface exists so a broker — and, in time, a customs authority — can see how the AI is doing its job, not just take its word. The methodology:
- Every broker correction becomes precedent. When a broker overrides a classification, that correction is captured with its reason, folded back into the eval set, and surfaced as a few-shot example on the next run. The loop is closed: the system that files tomorrow's entry has already learned from today's correction.
- Grounded, not guessed. Classifications are grounded on real HTSUS text and CBP rulings, and the model cites the ruling it relied on. That grounding is what produces measurably better accuracy than an ungrounded model — and 0% hallucinated codes: every code the AI proposes exists in the schedule.
- Calibrated confidence. The confidence score means what it says — the AI is honest about its own uncertainty, so the auto-approve threshold routes the genuinely hard lines to a broker instead of guessing.
- Reproducible eval harness. A CROSS-holdout harness re-scores the classifier on rulings it never saw, so accuracy and cost are measured, repeatable, and tracked over time rather than asserted.
| Surface | Status | Notes |
|---|
| Auto-approval rate headline | ✅ | Share of AI classifications that stood without a broker change |
| Override rate by chapter | ✅ | Names the chapters where the AI is weakest, so review effort goes where it's needed |
| Confidence calibration | ✅ | Shows whether stated confidence matches actual accuracy |
| Screening false-positive rate | ✅ | Per source list |
| Recent corrections feed | ✅ | What the AI just learned from broker overrides |
| Eval harness (CROSS-holdout) | ✅ | Reproducible; scores the classifier on held-out rulings |
ABI / CBP filing
| Feature | Status | Notes |
|---|
| Entry Summary (SE) message builder | 🟡 | Structural builder |
| AS2 transmission to CBP | ⬜ | Planned — gated on ACE certification |
| CBP response handling (A1/AE/CF28/CF29) | ⬜ | Planned |
Modules toggleable per deployment
The platform is a modulith — features can be enabled/disabled per deployment via COMMERS_MODULES. Reserved modules: supplier-network, post-entry-audit (in development).