Big week at Interfaze.

We got into Y Combinator P26. One years of research and building, from JigsawStack to Interfaze, and we have a lot more lined up for this year. We're going to share more product updates, challenges and thoughts. Wanted you to hear it first.

Here's what shipped this week.

40% faster first-token latency

Optimized the model serving layer. Most noticeable on longer documents and larger file inputs. Nothing to change on your end.

OCR improvement

Improved accuracy and bounding box precision on dense documents. If you're processing invoices, forms, or anything with tightly packed text, you should see better results.

Get started with document OCR: https://interfaze.ai/docs/vision/ocr

Speech-to-text response cleanup

Simplified the STT response structure. The speakers array is gone. Speaker labels are now a speaker field directly on each chunk. Same data, cleaner shape, roughly half the output tokens on transcription workloads.

Check out the latest STT: https://interfaze.ai/docs/audio/speech-to-text

If your app reads from speakers, update it to read from chunks instead.

Scraper improvements

Major speed improvement across most websites, around 10x faster on average. Also fixed a crash with certain selector inputs and improved metadata consistency.

Latest scraping engine: https://interfaze.ai/docs/web/web-scraping

Guardrails update

Tag format updated from GUARD to <guard> and added new categories:

ALL (Run all guard checks) S1_IMAGE Gore (Image) S15_IMAGE NSFW (Image)

Learn more here: https://interfaze.ai/docs/guardrails

That's it for this week!

Best, Yoeven