Hanji

building Hanji (YC X25). built @youlearnai for millions of students.

Hanji is an API for turning messy source documents (scans, faxes, handwritten forms, degraded government paperwork, dense tables) into structured data that downstream AI systems can actually use. The company is aimed at healthcare, insurance, and government teams , groups that tend to either maintain brittle in-house parsing pipelines or accept fixed tradeoffs between accuracy, cost, and latency from off-the-shelf providers. Hanji's pitch is that those three dials can be tuned per customer, with the vendor doing the tuning against the customer's actual document mix. The launch matters now because document extraction has become the bottleneck for a lot of AI applications in regulated industries, and generic OCR or single-model pipelines break on the exact edge cases (handwriting, faxes, broken tables) that these workflows depend on. The team frames Hanji around a playground you can try on real documents and a short onboarding where they tune the pipeline to a specific corpus, citing case work like lifting text accuracy from 71 percent to 96.5 percent for one healthcare customer, cutting end-to-end latency from 90 seconds to under a minute for another, and reducing processing cost by 75 percent for a third by adding a routing step. Hanji is built by Achyut Krishna Byanjankar, Soami Kapadia, and David Yu , part of Y Combinator's Spring 2025 batch. The founders previously built YouLearn, an AI tutor used by millions of students, where processing textbooks, handwritten notes, and scanned documents at scale is what pushed them to build the underlying parser that became Hanji.

Tags

Comments (7)