Nohena
A customs query lands days after you lodge, and the demurrage clock is already running. Nohena turns messy import documents into a correct, low-risk, ready-to-lodge declaration, with every output explained and a defensible audit trail, so you fix the problem before it costs you.
Nigerian import clearance is rigged against the agent. The customs authority scores every declaration with automated valuation and risk tooling, while the clearing agent usually finds out there is a problem only when the PAAR query and the demurrage clock arrive. Nohena flips that. It lets the agent see the declaration the way a customs risk engine will, before lodgement, and fix it while there is still time. This is pro-accuracy, not evasion. Nohena prepares; the licensed agent lodges.
How we took it on.
- 01
A modular monolith with a separate Python AI service behind an async job pattern, so it never blocks on a government portal or a provider outage.
- 02
RAG classification grounded in the tariff knowledge base (ECOWAS CET, WCO HS + explanatory notes, GRIs). The model never classifies from memory.
- 03
Rules-driven valuation (WTO transaction value + Article 8 additions) and AfCFTA / ECOWAS rules of origin, each decision citing the rule.
- 04
A two-checkpoint risk engine that predicts a four-lane outcome before lodgement, with strict per-tenant isolation and a tamper-evident audit trail.
What shipped.
The pipeline: vision-LLM extraction → RAG classification → valuation-risk → origin → checkpoint/lane, behind versioned contracts.
Four product surfaces: marketing, agent onboarding, the clearing console, and an internal review console.
NestJS API + Python AI service on Supabase / pgvector, with an NCS CET v3 + WCO HS knowledge base (~29.5k cited chunks).
Explainable by design: every classification, value flag, and origin call ships with a rationale, citation, and confidence.
The outcome, in plain terms.
Every decision carries a human-readable rationale, citation, and confidence. No naked verdicts.
Queries and demurrage caught early, with a two-checkpoint, four-lane risk view before lodgement.
A defensible, append-only audit trail that holds up under customs scrutiny.
The lodgement boundary respected end to end: Nohena prepares, the licensed agent files.
- NestJS
- Python
- Supabase / pgvector
- React + Vite
- RAG / Vision-LLM
- Vertex AI
- NDPA 2023
- Per-tenant isolation
- Append-only audit
- Eval-gated CI
Accuracy is the product. Every classification, value flag, and origin call ships with its rationale, citation, and confidence.