About Harmonova 01
WE REBUILT CUSTOMS DATA FROM THE SOURCE.
Harmonova is a real customs database. 928,552 national tariff lines across 149 national schedules, each traced to the official government schedule and revision it was parsed from. Not a model guessing at codes: a classifier that can only return a code the database already holds.
Harmonova HV-1 · power it on, watch it boot
Why we built it 02
Customs data deserved better.
Ten years moving goods across borders, watching the same failure: customs got harder, the data stayed poor, and every tool guessed. So we built the database we needed - sourced, versioned, and priced to the national line.
“Customs should be the easy part. The data just had to be right, and sourced, so people could trust it.”
How the database is built
Built from the official schedules, every line sourced.
Five principles behind the database. Every country covered, national depth across 149 schedules, and a source on every line.
- 01
Every product on earth, classified
The complete WCO Harmonized System. Every country, every category, no gaps.
5,635six-digit headings - 02
149 national schedules, seven of them complete to the line
928,552 national tariff lines. The US, Canada, Australia, New Zealand, Japan, the UK and the EU carried to full national-line depth, plus 142 further schedules at six-digit HS depth - every rate from the authority that sets it.
41,376US filable lines · 100%16,708CA filable lines · 100%16,605NZ filable lines · 100% - 03
A source on every line
Every rate carries the official schedule entry it was parsed from. Never an averaged rate.
- 04
Revision-stamped where the schedule moves fastest
EU, UK and US lines carry the published revision they were parsed from, so you can see which version your rate reflects.
- 05
A receipt on every lookup
Every result returns the code, rate, source and a confidence score, reproducible and audit-defensible.
Coverage 03
Every line, back to a sourced.
Every trading country on earth, covered by the complete WCO Harmonized System. 149 national schedules: 928,552 national tariff lines. Every rate parsed from the official schedule that published it.
- USITC HTSUnited States
- TARICEuropean Union
- Trade TariffUnited Kingdom
- WCO HSGlobal nomenclature
- World Bank WITS134 further schedules
Under the hood 04
Built to be audited, not taken on faith.
Take Harmonova past a security review or a technical evaluation. Here is exactly what it is, what it runs on, and why a language model on its own cannot replace it.
THE STACK
Purpose-built, and open where it counts
Purpose-built from a blank page, not bolted onto a decades-old tax suite. A TypeScript API over Postgres, deployed on Fly, where one schema unifies the complete WCO Harmonized System and 149 national schedules, and every line carries its source. Every line names the feed and schedule it was parsed from, so you can read the provenance behind a rate rather than take it on faith.
THE DATA PIPELINE
Parsed from the official schedule, sourced line by line
The database is sourced, not generated. Governments publish their tariff schedules as machine-readable data, not web pages. For the seven markets we go deepest on we ingest the authority's own published data: the USITC HTS export, EU TARIC, the UK Trade Tariff, CBSA, NZ Customs, ABF and Japan's MOF. The remaining 142 schedules come from the World Bank's WITS reference set, and every line records which of the two it came from, so you always know how close to the source you are standing. Everything is parsed into one normalised schema and stamped with the schedule it was parsed from. Nothing is guessed out of a rendered page, and a model never writes a tariff code, it only maps your product to one.
THE CLASSIFIER
Grounded, not guessing
Classification runs on Anthropic's Claude at temperature 0, wrapped in a deterministic category guard and a curated alias layer that catch the cross-chapter mistakes a raw model makes. Every output is schema-validated and checked against the headings the database actually holds. A code that is not in the schedule is never returned quietly: it is remapped to a valid line in the same heading, or, if the heading itself is bogus, surfaced with the confidence dropped and a specialist review demanded. It never invents a code and calls it confident.
SECURITY & PRIVACY
Built to pass a review
Postgres over TLS with certificate verification enforced, per-caller credentials granted route by route, and rate limits on every public path. Prompts are PII-redacted field by field before they leave our infrastructure, and we store the SHA-256 of a prompt for the audit trail, never its text. We do not train any model on your catalogue data. Our model provider's zero-retention terms are not yet contractually confirmed, and we will not claim them until they are. Data-residency deployment is on the roadmap, not shipped - ask us what is live today.
The question every technical buyer asks
“Can't anyone just build this with AI?”
Honestly? With months of work, maybe. But you would not be building an AI. You would be building a database.
Harmonova is not a model that guesses. It is a real customs database, built from the ground up, with every one of 928,552 national tariff lines linked back to the official government schedule it came from. 149 national schedules, with the seven core markets carried to full national-line depth. That is months of mapping, auditing and source verification, not a weekend of prompting.
Ask a chatbot for an HS code and it hands you a confident answer that fails an audit, because it never read the tariff schedule. A model can guess. Only sourced data and customs judgment can defend the answer.
The moat was never the model. It is the work, the sourced data, and the expertise behind every answer. We already did it, so your team does not have to.
Every duty line carries the schedule it came from
The established customs and tax suites, the Zonos and Avalara tier, are closed, priced for the enterprise, and rarely show you where a rate came from. Harmonova is API-native, API-native and built for a developer to drop in, and every single line traces back to its official government source. You are not trusting a vendor's black box. You are trusting the published schedule, with a link to prove it.