QNDN is our retrieval substrate — a Hopfield-inspired, asymmetric encoder/decoder whose measured edge lives in the deep tail of very large corpora, exactly where standard dense and lexical search structurally fail. One engine, trained once, generalises across biomedicine, patents, security, proteins, materials, and regulated archives — with zero per-vertical retraining.
Semantic encoders (BGE, ColBERT) map queries and documents into the same symmetric space. QNDN is deliberately asymmetric: a task-conditioned query head navigates a learned latent manifold toward documents via attractor dynamics — closer to a Hopfield associative memory than to cosine similarity.
The production stack fuses three complementary first-stage retrievers and reranks over their union. Each retriever has a different bias: a sharp head that lands obvious hits near rank 1, a shallow-tail head that reliably surfaces the non-obvious evidence ranked 800th that a human never scrolls to, and a lexical head for exact terms. Their union, reranked, is worth far more than any one alone.
Industry-standard self-hosted dense encoder. Sharp-head bias — lands the gold near rank 1 when it finds it.
Our 13M-param asymmetric query head over a 92M-param attractor-trained document encoder. Shallow-tail bias — gold reliably surfaces in the deep top-1000.
Classical exact-term retrieval. Anchors rare identifiers, codes, and names that embeddings smear together.
The same QNDN substrate plus an evidence-first reader generalises from legal text to patents to drug-repurposing atlases with zero per-vertical retraining. Swap the corpus, keep the geometry. Each vertical is named for what it does.
Grounded drug-repurposing and target discovery. Ask a research question; get a source-backed brief plus candidate molecules and proteins, every claim tied to its evidence.
Prior-art search and novelty / IP critique. Surfaces the non-obvious filings that anticipate a claim, and flags where a stack is crowded versus genuinely open.
Exploit-centric retrieval for bug-bounty and CTI. Turns a raw signal into "here's what that exact pattern led to in the wild" — disclosed reports, audit findings, before/after-fix pairs.
Asymmetric retrieval over protein space: sequence-only → candidate structural cousins, no structure required up front. The same Hopfield geometry, applied to biology.
Evidence-grounded materials scouting — e.g. thermal-interface and accelerator materials. Pairs the discovery engine with prior-art checks before you ever touch a lab.
Evidence-first retrieval for corpora you can't send anywhere. Air-gapped, calibrated refusal as a first-class output, one commodity GPU, no external APIs.
We don't beat SOTA dense retrievers on small-pool top-1. The measured win is where it matters for discovery and regulated search: scaling decay, complementarity under rerank, and calibrated refusal. Numbers below are from our internal scaling protocol, not a buyer corpus.
Single-retriever P@1 at 3.52M is 0.068; BGE-large is 0.149. The advantage is in scaling decay, complementarity under rerank, and calibrated refusal — not native P@1 on a small pool. The honest comparison is on a buyer's own corpus, not on this page.
The lab's operating mode isn't a publication pipeline; it's a measurement pipeline. Experiments run nightly, results go into a single versioned journal, and the product inherits only what survives ablation.
Researchers and teams interested in Hopfield-on-natural-language, retrieval scaling, asymmetric protein/biomedical retrieval, or a domain pilot can write to the lab directly. We share reproducible pipelines for any number on this site, typically under NDA or academic-collaboration agreement.
Response time: typically 2–5 business days.