Document Cleaner was built around a plain truth of insurance operations: the documents are never as clean as the workflow wishes they were. The public-safe version lives at UwUGreed/Doc-Cleaner.
Submissions, exports, PDFs, attachments, notes, and supporting material tend to arrive in inconsistent shapes. Some files are too large. Some are mostly noise. Some are useful only after the right sections are pulled forward. If you send that mess directly into an AI workflow, the model spends too much context on junk and the user spends too much time sorting out what happened.
The goal of Document Cleaner was to sit before the AI step and make the intake material easier to reason about.
The Problem
Most AI demos quietly assume the input is already nice. Real business documents are not like that.
In the insurance workflow, preparation mattered as much as generation. A cleaner packet made it easier to:
- Review submissions faster.
- Keep supporting material together.
- Reduce repeated manual cleanup.
- Pass better context into local AI tools.
- Avoid burying important details under formatting noise.
The system was not meant to replace review. It was meant to make review less miserable.
The Build
The useful part of the project was the pipeline shape.
Document Cleaner treated document prep as a sequence of small, bounded transformations rather than one giant "AI clean this" prompt. That made the behavior easier to inspect and easier to control.
The workflow focused on:
- Normalizing messy intake material into a more predictable structure.
- Reducing low-value formatting and duplicate noise.
- Preparing smaller, more focused context packets for AI-assisted review.
- Keeping the output readable enough for a human to verify.
- Staying inside the local workflow boundary instead of becoming an internet upload tool.
That last part mattered. The project existed in a sensitive business environment. The right architecture was not "send everything to the most convenient API." The right architecture was local, deliberate, and reviewable.
What Made It Different
The design challenge was not just extraction. It was restraint.
If a document tool rewrites too aggressively, it can create false confidence. If it leaves everything untouched, it does not save enough time. The middle ground was to clean and organize the material while keeping the human close to the final decision.
That meant the output needed to be useful, but not magical. It needed to support the person working the account, not pretend the system had full business authority.
Why It Matters
This project changed how I think about applied AI.
The model is rarely the whole solution. The work around the model matters just as much: document hygiene, file boundaries, context shaping, review loops, and operational trust.
Document Cleaner was one of the projects that made that obvious. Before AI can help a workflow, the workflow has to know what it is handing to the AI.