Concord: freight documents that agree before customs
Customs brokers and freight forwarders in Baku handling Middle Corridor shipments. Every shipment comes with an invoice, packing list, certificate of origin and air waybill, filled in by different parties. Today an employee cross-checks them by eye: 5-20 minutes per shipment (interview: Air Cargo Azerbaijan). The most frequent errors are in consignee details and in amounts, for example a 5 USD difference between line items and the invoice total. The same company appears in Azerbaijani, English and Russian spellings, so simple matching fails. Concord reads all four documents, cross-checks them, shows every discrepancy with the value and location from each document, and drafts the correction email to the shipper.
Preliminary score
- Value for the user22.0 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI25.0 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing18.0 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility11.5 / 15
Data requirements, running costs and a clear next step.
- Originality7.5 / 10
A meaningful difference from common solutions.
A finalist: the live jury scored it 63.0 / 100 on stage.
What the judges wrote
- GPT-6 AstraPreliminary83/100
- Strengths
- Your industry interview grounds a specific, measurable workflow problem: brokers spending 5–20 minutes cross-checking shipment documents. Your separation of AI extraction and multilingual name matching from deterministic arithmetic checks is well suited to the task, with evidence attached to findings. Your reproducible holdout evaluation, image-quality comparisons and candid failure analysis provide substantially stronger testing evidence than generic accuracy claims.
- To improve
- Your first priority should be making unreadable fields require human review, with regression tests preventing silent passes. Validate your gains on real shipments against timed broker checks and a transliteration-aware baseline, measuring total handling time rather than pipeline latency alone. Include an input-to-email execution trace and reconcile the deck's Claude references with the GPT-6 Luna evaluation configuration. Quantify API cost per shipment and document provider-side retention controls; a forwarder's own API key does not by itself establish confidentiality.
- Claude Opus 5.5Preliminary85/100
- Strengths
- Your problem is concrete and grounded. You name a specific user (declarants at Baku forwarders), a measured pain (5–20 minutes per shipment from a same-day industry interview), and the exact error types brokers report. Your split between the LLM (extraction, cross-script company matching) and deterministic code (weights, totals, VÖEN, HS codes, the email) is well reasoned, and the UI shows which one decided each finding. Your evaluation is unusually honest: a held-out seed, an exact-match baseline, three input qualities, per-type results, and failures stated plainly, including recall collapsing on bad phone photos and the silent-pass bug.
- To improve
- Put a currency figure on running cost (for example dollars per shipment at current API prices) rather than only call count and latency, so a forwarder can weigh it against staff minutes. Your test set is small (15 shipments, 14 errors) and fully synthetic with errors you injected yourselves, so a few real anonymised shipments would greatly strengthen the precision claims. Implement the 'Yoxlanmalı instead of silent pass' fix you describe, since silent misses are the riskiest failure for a compliance tool. Make the deck consistent on the model: slides 3–4 say Claude, while your disclosure and test results cite GPT-6 Luna.
The team
Rahman Pashayev
business
Gasim Garayev
product
Ravan HajızadaLead
developer