Feyn Labs says Pulpie cleans web pages for AI training at one twentieth Dripper's cost

Why it matters

Web extraction is a cost center inside AI pipelines. Feyn is betting that a cheap, task-specific encoder can make cleaner training and retrieval data affordable outside frontier labs.

Feyn Labs co-founders Shreyash Nigam (@shreyash_nm) and Bhavnick Minhas have released Pulpie, a family of HTML content-extraction models built to strip navigation, ads, sidebars, footers, and other boilerplate from web pages before they are used in AI training or retrieval systems.

The San Francisco company published the technical field note on June 25th and pushed the pulpie 0.0.2 package to PyPI on July 1st. The release puts Feyn's commercial thesis into code: Nigam's founder memo says the company wants customers to own task-specific models trained on their data, while Feyn's Y Combinator profile describes an end-to-end service for building custom models that keep improving from production feedback.

Nigam, Feyn's founder and CEO, previously worked on on-device AI, ads personalization, and VM memory at Google, according to YC. Minhas, Feyn's CTO and co-founder, is listed by YC as building document AI support for Chonkie. The two had already built Chonkie, an ingestion and chunking library for RAG pipelines that now shows 4.4k GitHub stars. Pulpie moves the team upstream from chunking text to deciding which text should exist in the pipeline at all.

The claim: Dripper-like quality at a fraction of the cost

Feyn says its recommended model, pulpie-orange-small, has 210 million parameters and scores 0.862 ROUGE-5 F1 on WebMainBench. That nearly matches Dripper, which Feyn reports at 0.864 with a 600 million parameter decoder. The benchmark is company-run, and Feyn's field note does not show independent reproduction, so the numbers should be read as Feyn's measurements rather than market-validated performance.

The cost claim is sharper because it depends on throughput. Feyn says pulpie-orange-small processes 13.7 pages per second on an NVIDIA (@nvidia) L4 GPU, compared with 0.68 pages per second for Dripper on the same hardware. At an L4 price assumption of $0.39 per hour, Feyn calculates the cost of cleaning 1 billion pages at about $7,900 with Pulpie Small versus about $159,000 with Dripper.

The reason, according to Feyn, is architecture. Dripper is a decoder-style extractor that emits labels step by step. Pulpie uses an encoder that labels every HTML block as content or boilerplate in one forward pass. That matters on cheaper GPUs: Feyn argues that decoder extraction is bound by memory bandwidth, while its encoder approach is closer to a dense compute problem. On the A100, Feyn says Pulpie Small runs 7.1x faster than Dripper; on the L4, the gap widens to 20x.

The model family has three sizes: Pulpie Orange Small at 210 million parameters, Pulpie Orange Base at 610 million parameters, and Pulpie Orange Large at 2.1 billion parameters. Feyn says the large model is the teacher, with the smaller two distilled from it. On WebMainBench's English subset of 6,647 pages, Feyn reports 0.873 ROUGE-5 F1 for Large, 0.863 for Base, and 0.862 for Small.

What Pulpie actually does

The PyPI package wraps a four-stage pipeline. Pulpie first simplifies raw HTML by removing scripts, styles, and formatting noise, then tags each block with an ID. It chunks the page into sequences of up to 8,192 tokens, classifies each block as main content or boilerplate, and returns the kept blocks as HTML or Markdown.

The default install path is simple:

pip install pulpie

Then:

from pulpie import Extractor

extractor = Extractor()
result = extractor.extract(html)
print(result.markdown)

That simplicity is the point. AI companies already spend heavily on crawling, filtering, deduplication, and retrieval. A low-cost extractor that preserves tables, formulas, code blocks, and article text would cut waste in both pre-training corpora and retrieval-augmented generation systems. Feyn's field note says that during discovery, it found about 70% of blocks on a typical HTML page were boilerplate.

Pulpie's training set also shows the practical constraint behind the work. Feyn says there was no public block-labeled HTML dataset large enough for its needs, so it sampled 16,670 English pages from Common Crawl, limited the sample to one page per domain, and used DeepSeek V3.2 to label blocks. After filtering, Feyn says it kept 15,880 pages, then ran Dripper 0.6B as a second labeler and retained 14,959 pages where the two labelers agreed on at least 70% of blocks.

The licensing detail commercial teams should read

Feyn calls the Pulpie models open source in the field note and hosts the family on Hugging Face (@huggingface). The implementation story is split across licenses. The Pulpie GitHub repository and PyPI package list Apache 2.0. The Hugging Face card for the default Small model, however, lists the weights under CC BY-NC 4.0 and says to contact Feyn for commercial licensing. The Base and Large model cards list Apache 2.0.

That distinction matters because Feyn is not only publishing a research artifact. Feyn is selling a broader model-building service, and Pulpie is a proof point for the same argument Nigam makes on the company homepage: generic models are increasingly abundant, but the advantage shifts to better task-specific data systems and owned models. Pulpie is a small, concrete version of that pitch. Clean the page before it enters the model, and the rest of the AI stack has less garbage to explain away.

For developers, Pulpie competes with older structure-based extractors such as Trafilatura, Readability, magic-html, and Boilerpipe, plus newer model-based parsers like Dripper. For Feyn, the release is also distribution. A working open package gives the startup a way into engineering teams that already feel the cost of messy web data, then points those teams toward Feyn's paid custom-model work when generic extraction is not enough.