TraceLot

Inspiration

When a supplier reports a quality issue or changes a material specification, the answer to “Which products could be affected?” may be scattered across purchase orders, inspection reports, assembly records, and shipment documents. Manually connecting those records takes time, and missing information can matter as much as a possible match.

I built TraceLot to make that investigation faster and easier to verify.

What it does

TraceLot takes a supplier alert and searches supporting records for matching component and supplier-lot identifiers. It displays the matching source records, an investigation timeline, missing information, and a draft checklist for a quality reviewer.

If shipment or finished-product records are missing, TraceLot flags the gap instead of guessing. It is an investigation assistant: it does not declare a product defective, initiate a recall, quarantine inventory, or contact customers.

How we built it

TraceLot uses a Python FastAPI backend and a browser-based dashboard. Its demo mode runs repeatable investigations over fictional supplier records. Its live mode can read supported documents from a configured Box folder, extract text from PDFs, DOCX files, and text files, and use Amazon Bedrock to draft a cautious investigation summary and reviewer checklist.

Exact identifier matching keeps the evidence trail tied to source filenames and the identifiers found in the records. The generated summary supports the review; it does not replace the source evidence or the quality team's decision.

Challenges we ran into

The key challenge was making an AI assistant useful without letting it turn incomplete records into a confident operational decision. We designed TraceLot to show its source records, flag missing evidence, and leave consequential decisions with a human reviewer.

We also wanted the demo to work before cloud services were configured, so we included fictional sample scenarios alongside the Box and Bedrock integration path.

What we learned

In document-heavy workflows, a useful agent needs to do more than summarize. It should show how it found a possible connection, identify what it could not verify, and help a person decide what to check next.

What's next for TraceLot

We would add source-page citations for PDFs, richer lot relationship extraction, and a reviewer-controlled case export. Any production use would need validation with quality experts and the organization's procedures.

Built With

• agentic-ai
• amazon-web-services
• bedrock
• box
• document-ai
• fastapi
• manufacturing
• python
• quality-management
• supply-chain