Hey everyone 👋, we’re David, Soami, and Achyut from Hanji.

At YouLearn, we processed 70M pages for 2.5M users. To handle 500 page textbooks, handwritten notes, and scanned documents with broken tables, we built our own document parser. We launched it as an API and didn’t expect healthcare, government, and insurance teams to be the first to reach out. But after talking with them, we realized they still couldn’t find a pipeline they could rely on in production. 

https://youtu.be/S5OVXZVv61A

Problem

As companies process more documents with hard formats, they either build in-house and take on the cost of maintaining it, or use a provider and accept tradeoffs they cannot control.

• Build: It can work at a certain scale, but as volume increases and harder formats come in, your team has to keep fixing the pipeline. Over time, you end up maintaining something that was never your product.
• Buy: Providers can work too, but you have to accept tradeoffs you cannot change. They may cost too much, take too long at higher volumes, or miss important cases that your team then has to fix.

In both scenarios, your team pays for these tradeoffs through more engineering time, slower workflows, or lower margins.

Our Solution

Hanji tunes the document pipeline to your workflow, so you don’t have to maintain it yourself or accept fixed tradeoffs.

Send us your hardest docs, and we’ll take care of the rest by tuning your pipeline to the accuracy, speed, and cost you need. And when human review is still necessary in your document workflow, we’ll improve that process too, such as by adding field-level confidence scores and review flags.

Benchmark

Even before tuning, Hanji holds the highest text accuracy across five providers on our benchmark: 400 human-verified pages spanning seven document types.

Text accuracy, 400 human-verified pages, 95% CIs. Full methodology here.

Results

After tuning Hanji to each customer’s document workflow, one YC healthcare company improved their text accuracy from 71% on our first pass to 96.5% within a few days. Another reduced its end-to-end latency from 1.5 minutes to under 1 minute with Hanji extracting 78/78 fields containing payer to member ID associations across 53 packets. And another company reduced its projected processing costs by 75% after we implemented a routing step that found the relevant forms in every packet.

In each case, we started with their own documents, tuned the pipeline around what matters for their workflow, and kept improving it as new failures appeared.

Our Ask

Try your hardest docs in the playground here, then book a call to see how we can improve your accuracy, speed, or review process.

Processing regulated documents? Book a call. For PHI documents, we can sign a BAA and enable access the same day.

For any other questions or feedback, email us directly at founders@hanji.dev.