Papers, benchmarks, evals, and long-form technical posts from the neww.ai team. We publish what we learn — successes and failures.
We describe the engine-registry pattern used at neww.ai to unify 53+ AI engines behind a single routing layer, and present empirical results on end-to-end latency and quality across 30 downstream products.
ReadWhat we learned running RAG across millions of documents — chunking, hybrid retrieval, re-ranker cost-benefit, and eval harness design.
ReadA public benchmark evaluating model selection policies across cost, latency, and quality for production tool-use workloads.
An evaluation of prompt-injection defenses when agents chain across untrusted inputs and tools, with recommendations for practitioners.
We partner with academic groups and independent researchers on evals, benchmarks, and open problems.
research@neww.ai