Launched this week
KYNEX Harness
Agentic Extraction harness that beats GPT-6 Astra and Opus.
2 followers
Agentic Extraction harness that beats GPT-6 Astra and Opus.
2 followers
KYNEX gives a language model six tools — page text, table parsing, OCR, vision, full-text search, and layout analysis — and lets it decide what to inspect. No chunking, no embeddings, no retrieval step. Scores 92.97 F1 on ExtractBench (370 documents, 67 types) using GPT-5.4, above GPT-6 Astra (91.9) and Claude Code (87.1). Verified by LlamaIndex. Apache-2.0.
Free
Launch Team / Built With
Render: Get up to $100K in startup creditsThe fastest way to launch and scale your startup in the cloud
Promoted
I built KYNEX because I noticed a pattern: bigger models weren't solving structured extraction. The bottleneck wasn't reasoning — it was document access.
A 26,725-row schedule can't fit in one LLM response, so the model returns null. A scanned page has no text layer, so it's invisible. A table where the model reads values "by eye" drifts on separators and symbols.
Instead of prompting harder, I gave the model tools and let it decide what to inspect. Five post-processing rules — page-wise enumeration, non-record pruning, redaction markers, structured table rows, OCR fallback — each improved specific documents from near-zero to 0.83-1.00 F1. Every rule has an individual kill switch so I could measure its effect.
The result on ExtractBench: GPT-5.4 + tools scores higher than GPT-6 Astra without them. I also use this at ABB for RFQ summaries and CPQ configuration sheets.
Honest limitation: the rules were developed on the same benchmark documents, and long documents still score lower (78.77 vs 94.29 on short). I'd like feedback on how to evaluate post-processing without test-set leakage.