Building the design foundation for a document-intelligence platform.
Reducto is a document ingestion API used by infra teams who need documents read exactly right. The output is dense: bounding boxes, confidence scores, field level extractions, schema edits.
My job was the layer where that becomes legible, so a developer can see what the model saw, spot where it went wrong, and fix it.
Structuring unstructured data is reducto’s bread and butter. One way to do it is by setting up extract. Here you can define a schema that takes all the data from a certain document and makes it database ready. However, defining each field can be quite challenging. Schema optimizer was a tool I built to give them best practices guidance quickly.
After you’ve optimized, sometimes it takes a bit more granular testing to get to 100% accuracy. The evaluation tool runs a costly model once to set ground truths, then tweaks your cheaper pipeline settings to get it to that 100%. This keeps accuracy high and price low making their pipeline runs more scalable.
Redaction is a huge issue for many companies. Typically managed by in house teams and can be an incredibly tedious process. Furthermore, since it’s usually done by humans, they can miss things such as removing it from the metadata too. Redaction was a new feature I proposed to the team.
