What we do
Research papers, engineering and opinion blogs, the prior work this company is built on, and open benchmarks you can run yourself. Every claim links to where it is published.
Research papers
Peer-visible work with numbers you can check. Papers that come in through the publication pipeline appear here automatically.
Blogs
Engineering deep-dives and business perspectives from the team building Kurious.
Blended RAG: Improving RAG Accuracy with Semantic Search
94.89% Top-5 accuracy on SQuAD without fine-tuning. The paper, explained.
Read →Finding a Needle in 4,000 Hours of Legal Video
From single-pass retrieval to agentic deep thinking. Quick mode: 6.9s, 90% accuracy. Deep mode: 94%.
Read →Your AI Has Never Seen Your Data
85 million records. 23 agencies. 8 formats. Every frontier model sees none of it. Here is why, and what we built.
Read →Prior research
The results this company is built on. Each one links to where it was published or measured.
Blended RAG
SQuAD Top-5. No fine-tuning. No frontier model.
arXivNJ Open Data
85M records, 23 agencies. Multi-format government benchmark.
HuggingFace leaderboardChemRAG
Chemical domain retrieval. 2.4M reactions indexed, retrosynthetic pathways.
HuggingFace leaderboardLegal Video
4,000+ hours. Deep mode. Video moment retrieval.
HuggingFace leaderboardBuilt by the minds behind 50+ patents and 80+ research papers.
Open benchmarks on HuggingFace
Verify our results. Run your own models against the same benchmarks.
All publications
Everything that has come through the publication pipeline: papers, technical write-ups, customer stories, videos and off-site pieces. Newest first.