We train the specialists production AI runs on.
ZeroEntropy is a specialized training and inference company. We build the small, calibrated models that production AI systems read scalars off of — rerankers, embedders, classifiers, judges, faithfulness scorers, and the long tail of scalar axes a real product needs measured. The bet underneath the company is that the long-term shape of production AI is a constellation of fine-tuned specialists wrapped around frontier LLMs, not one general-purpose model doing everything.
The generalist race writes the headlines. Specialists ship the systems. A 0.5B reranker calibrated against a customer's actual eval harness beats a frontier LLM on the metric that matters at roughly 1% of the cost and 10% of the latency. The same shape holds for embedding models, hallucination judges, instruction-adherence classifiers, prompt-injection detectors, and any other narrow scalar a production stack reads off a model. Frontier LLMs are the wrong unit of compute for that work — they pay for breadth no one is using on a per-call basis.
The capability behind those models is the moat — not any individual model. A calibrated universal pairwise comparator over an axis of comparison gets distilled into a small pointwise scorer that ships at inference scale. Relevance was the first axis we shipped at production quality; that is what zerank-2 and zembed-1 are. The same pipeline extends to any other axis a customer cares about — faithfulness, hedging, instruction adherence, tone, prompt-injection risk, code-bug probability, factuality, scientific rigor. Each new axis is another permanent artifact: an open-ended gallery of specialists trained the same way.
Production AI is a constellation of specialists wrapped around frontier LLMs. Not one big model doing everything.
The team is small and dense. Competitive-math fluency on the bench, ex-quant discipline about numerical correctness and the cost of a basis point of latency, founder-grade obsession with shipping. There is no separate research arm. The engineer who proves the training objective is the engineer who runs it against the customer's corpus the next day. This page names no one on purpose — every model, every eval, every catalog entry is collective output of that bench.
What ships out of ZeroEntropy: the model line (zerank-2, zembed-1, and the specialists that follow); the training methodology behind them (zELO, chess-Elo turned into a continuous relevance training signal); the public eval harness we hold our own models to (MTEB Evals); and the catalogs of how the field actually works in production — Concepts, Playbooks, Versus. The catalogs are not marketing. They are the densest possible signal of what we know, in the cheapest possible format, with the longest possible half-life.
Production AI built on specialists is the conversation: contact@zeroentropy.dev. The bench hires for olympiad-grade fundamentals and shipping obsession: careers@zeroentropy.dev.
