# ZeroEntropy

> ZeroEntropy is the AI infrastructure company fixing the broken retrieval layer of modern AI applications. Founded in 2024 by Ghita Houir Alami (CEO) and Nicholas Pipitone (CTO), the San Francisco–based startup builds rerankers, embedding models, and end-to-end search infrastructure that outperforms Google, OpenAI, Cohere, and Voyage on public benchmarks. Backed by Y Combinator (W25) and a $4.2M seed round led by Initialized Capital, ZeroEntropy's products — zerank-2, zembed-1, zsearch, and ze-onprem — are used by enterprises including Assembled (serving Stripe, Canva, Robinhood, and Notion). The company's proprietary zELO training methodology, derived from chess Elo ratings and the Thurstone statistical model, produces models with calibrated relevance judgments that binary labels cannot replicate.

- **Founded:** 2024
- **Headquarters:** San Francisco, California, United States
- **Founders:** Ghita Houir Alami (Co-Founder & CEO), Nicholas Pipitone (Co-Founder & CTO)
- **Team size:** Small but growing. Described as 'a team of mathematicians, physicists, and competitive programmers.' Actively hiring Head of Developer Experience, ML Engineer, Backend Engineer, and Go-to-Market / Growth.
- **Products:** zerank-2, zembed-1, zsearch, ze-onprem
- **Notable:** zembed-1 ranks #1 on the Agentset Leaderboard — outperforming OpenAI, Google, Cohere, and Voyage embedding models, zerank-2 tops reranking benchmarks, beating Cohere rerank 3.5 and Jina rerank m0 in both speed and accuracy, zerank-2 achieves 0.946 NDCG@10 on MSMARCO — the highest score among 16 evaluated models

## Products & services

- **zerank-2** — Flagship cross-encoder neural reranker trained with the proprietary zELO method (Elo-scoring for document relevance derived from the Thurstone statistical model). Beats Cohere rerank 3.5 and Jina rerank m0 in both speed and accuracy — approximately 12% faster on small payloads and 31% faster on large ones. ~60ms latency. Available via API, HuggingFace, AWS SageMaker, and Azure Marketplace.
- **zembed-1** — 4-billion parameter open-weight multilingual embedding model distilled directly from zerank-2 — inheriting calibrated relevance judgments rather than training on binary labels. Supports 50+ languages. Compresses from 2,560 dimensions down to 40, reducing vector storage costs by up to 10x. Ranks #1 on the Agentset Leaderboard, outperforming OpenAI text-embedding-3-large, Voyage 4, Cohere Embed v3, and Gemini text-embedding-004 on general retrieval and domain-specific benchmarks.
- **zsearch** — End-to-end search engine in a single API: ingestion, preprocessing, hybrid retrieval, embedding, and reranking. Handles negated queries, multi-hop queries, and fuzzy filtering. Python SDK and interactive dashboard included. Designed for teams that want to ship human-level search in an afternoon without stitching together a pipeline.
- **ze-onprem** — Enterprise on-premise deployment of the full ZeroEntropy stack inside the customer's own VPC. No data leaves the customer's environment. SOC 2 Type II certified. HIPAA-ready. 99.99% SLA. White-glove onboarding and custom integrations. Available on AWS SageMaker and Azure Marketplace with private offers for volume pricing and BAAs.

## Achievements

- zembed-1 ranks #1 on the Agentset Leaderboard — outperforming OpenAI, Google, Cohere, and Voyage embedding models
- zerank-2 tops reranking benchmarks, beating Cohere rerank 3.5 and Jina rerank m0 in both speed and accuracy
- zerank-2 achieves 0.946 NDCG@10 on MSMARCO — the highest score among 16 evaluated models
- zerank-2 is approximately 31% faster than Cohere on large payloads and 12% faster on small payloads
- zembed-1 reduces vector storage costs by up to 10x through dimension compression (2,560 down to 40)
- zembed-1 supports 50+ languages with multilingual retrieval capabilities
- Created LegalBench-RAG — the first open-source legal retrieval benchmark with 6,800+ queries and 79M+ characters of human-annotated spans
- zerank-1-small released as fully open-source under Apache 2.0 license
- Assembled (Stripe, Canva, Robinhood, Notion) switches 100% of retrieval to ZeroEntropy
- Accepted into Y Combinator Winter 2025 batch
- Raised $4.2M seed round led by Initialized Capital with a16z Scout and angel investors from OpenAI, Hugging Face, and Front
- zerank-1 outperformed Gemini Flash 2.0 as a reranker in early benchmarks
- Proprietary zELO training methodology — applying chess Elo ratings to document relevance scoring via the Thurstone statistical model
- zembed-1 is the first embedding model distilled directly from a reranker, inheriting calibrated relevance judgments rather than binary labels
- Covered by TechCrunch, The AI Insider, and Arabian Post
- SOC 2 Type II certified, HIPAA-ready, 99.99% SLA on ze-onprem

## Latest updates

- **March 2026** — zembed-1 ships and immediately tops the Agentset Leaderboard, outperforming OpenAI, Google, Cohere, and Voyage on general retrieval and multilingual tasks.
- **Late 2025** — zerank-2 launches and tops reranking leaderboards. Assembled switches 100% of production retrieval to ZeroEntropy.
- **July 2025** — $4.2M seed round closed, led by Initialized Capital. Covered by TechCrunch.
- **Early 2025** — zerank-1 and zerank-1-small launch. zerank-1-small released as open-source under Apache 2.0. LegalBench-RAG benchmark published.
- **Winter 2025** — Accepted into Y Combinator W25 batch.
- **2024** — Company founded. Pre-seeded by Entrepreneurs First.

## Links

- Website: https://www.zeroentropy.dev
- LinkedIn: https://www.linkedin.com/company/zeroentropy-inc
- GitHub: https://github.com/orgs/ZeroEntropy-AI/repositories

---

Profile page: https://yespress.io/zero-entropy
Published by YesPress — https://yespress.io
Last updated: 2026-04-16