CargoLens AI

Inspiration

CargoLens AI was inspired by the challenge customs and cargo officers face when inspecting unfamiliar products. A declaration may use a broad or vague description, while the physical item, packaging, label, model, or specification tells a more detailed story. Officers cannot be experts in every product category, so they need faster access to evidence and practical next steps.

What it does

CargoLens AI is an inspection copilot that compares cargo images with invoices and declaration documents. It identifies visible product attributes such as brand, packaging, product type, variant, volume, and specification. It then classifies findings as a match, partial match, mismatch, or insufficient evidence. The app is intentionally human-in-the-loop. It does not make final customs, duty, clearance, quality, purity, grade, or authenticity decisions. Instead, it explains what is visible, what cannot be verified, and what the officer should inspect next.

How we built it

We built CargoLens AI with FastAPI, Jinja2, Bootstrap, JavaScript, SQLite, SQLAlchemy, and ReportLab. Codex with GPT-5.6 accelerated the development of the frontend, backend architecture, upload flow, inspection history, structured analysis, report generation, and project organization. The application has a provider-based AI layer. It supports an OpenAI-compatible analysis path and a local Ollama fallback. For local testing without an API key, we used Ollama with Qwen2.5-VL to analyze cargo images.

Challenges we ran into

The main challenge was handling multimodal analysis locally. Large cargo photos and invoice scans can exceed the local model context limit. We also found that a model can be overly critical when information is missing. We refined the prompts and risk logic so that unreadable labels, invisible material, unknown quantity, or unverified grade are treated as insufficient evidence—not automatically as a mismatch. Another challenge was handling multiple invoice items and multiple cargo images. This led us to move toward item-by-item analysis instead of one broad shipment-level comparison.

Accomplishments that we're proud of

We are proud that CargoLens AI became a complete inspection workflow rather than only an image classifier. It includes: Cargo and document uploads AI-supported product identification Risk-aware inspection results Evidence and limitation reporting Recommended verification actions Inspection history Follow-up chat Downloadable inspection reports A conservative human-in-the-loop design We are also proud that the project can run locally with a fallback model, making the prototype accessible without requiring paid API access.

What we learned

We learned that product recognition alone is not enough for an inspection workflow. The useful part is reasoning across physical evidence, documents, uncertainty, and officer actions. We also learned that: Visible branding does not prove authenticity. Photos cannot reliably confirm grade, purity, quality, duty classification, or legal compliance. Invoice extraction must be separate from cargo-image analysis. Missing evidence should not be presented as fraud. Clear explanations and limitations are as important as confidence scores.

What's next for CargoLens AI

Next, we plan to add dedicated invoice-image extraction, OCR support, and robust multiple-item matching. We also want to add: Item-level comparison tables Barcode and QR-code scanning Manufacturer catalogue/reference-image comparison Trusted product-data integrations Better invoice and packing-list parsing Role-based officer workflows Audit logs and secure cloud storage Deployment with real GPT-5.6 multimodal analysis when API access is available

Built With

• bootstrap
• codex
• css3
• cv
• fastapi
• gpt-5.6
• html5
• javascript
• jinja
• ollama
• python
• qwen2.5-vl
• sqlalchemy
• sqlite