BREAKING

⚡ ZeroEntropy raises $4.2M seed — Initialized Capital leads, YC follows ◆ 🏆 zerank-2 tops reranker benchmarks — beats Cohere, Jina, Salesforce ◆ 🚀 zembed-1 ranks #1 on Agentset Leaderboard — outperforms OpenAI, Google, Voyage ◆ 🔍 Assembled (serving Stripe, Canva, Robinhood) switches 100% of retrieval to ZeroEntropy ◆ 📰 Featured in TechCrunch · The AI Insider · Arabian Post ◆ ⚡ ZeroEntropy raises $4.2M seed — Initialized Capital leads, YC follows ◆ 🏆 zerank-2 tops reranker benchmarks — beats Cohere, Jina, Salesforce ◆ 🚀 zembed-1 ranks #1 on Agentset Leaderboard — outperforms OpenAI, Google, Voyage ◆ 🔍 Assembled (serving Stripe, Canva, Robinhood) switches 100% of retrieval to ZeroEntropy ◆ 📰 Featured in TechCrunch · The AI Insider · Arabian Post

YesPress. Company Profile

ZeroEntropy

"The Supabase for Search." — CEO Ghita Houir Alami

Most AI products lie to you — not because the model is wrong, but because it retrieved the wrong thing. ZeroEntropy fixes that. Their rerankers, embeddings, and search infrastructure are the unsexy, load-bearing wall that keeps AI from hallucinating its way into your legal contract or medical record.

YC W25 $4.2M Raised San Francisco Founded 2024

Visit ZeroEntropy →

#1 Agentset
Leaderboard

The Setup

AI doesn't hallucinate.
It retrieves badly.

Three panels. One uncomfortable truth.

01 🧩

The Cobbled Mess

Most teams stitch together a vector database, a keyword search, and a re-ranker from three different vendors. Then maintain it. Then watch it break at 2am.

02 😤

The Context Dump

Others dump the entire knowledge base into the LLM context window. Quick fix. Compounding errors. Expensive. The model confidently answers the wrong question.

03 ⚡

The ZeroEntropy Way

One API. Ingestion, indexing, reranking, evaluation. Deploy human-level search in an afternoon. Fewer hallucinations. Faster results. Actually works.

The Founders

Two brains.
Infinite equations.

A Moroccan mathematician who left home at 17, and a CMU dropout who audited blockchains for hedge funds. Somehow, this works.

Ghita Houir Alami

CEO · The Architect

Born in Morocco, Ghita left at 17 to study at École Polytechnique in Paris — France's most elite mathematics and engineering institution. She followed that with a master's in Applied Mathematics at UC Berkeley, where the idea for ZeroEntropy quietly took shape. Before ChatGPT went viral, she was already building conversational AI — and hitting the retrieval wall every time. That frustration became a company.

At 25, she is one of the very few female CEOs building deep AI infrastructure. Featured in TechCrunch and The AI Insider. Actively inspiring young women across Morocco and Africa to pursue engineering careers.

École Polytechnique · UC Berkeley · TechCrunch

Nicholas Pipitone

CTO · The Builder

Grew up doing math and coding competitions. Dropped out of CMU to build startups. Became CTO at five different companies. Wrote low-level C, C++, Assembly and GPU code. Built stat-arb algorithms. Audited blockchains for bug bounties and hedge funds. At some point he built the AI at Myko.ai, Manifestapp and MagiBook.

The kind of engineer who treats a GPU as a musical instrument. Brought to ZeroEntropy the deep systems expertise that made zerank and zembed possible.

CMU · 5x CTO · Competitive Coder

"In information retrieval, entropy refers to the variability of information in a message. ZeroEntropy refers to knowing exactly what the message contains." — The name, explained

Origin Story

Pre-seeded by Entrepreneurs First.
Polished by Y Combinator.

ZeroEntropy's earliest backer was Entrepreneurs First, which pre-seeded the company before the YC application. Both founders had spent years inside AI projects — in healthcare, finance, B2B SaaS, and consumer apps — encountering the same retrieval failures again and again. Every team was reinventing the same broken pipeline. They decided to fix it once, for everyone.

Y Combinator's Winter 2025 batch took them in. The $4.2M seed round that followed was led by Initialized Capital, with participation from Transpose Platform, 22 Ventures, a16z Scout, and angels from OpenAI, Hugging Face, and Front.

"Everyone was reinventing
the same broken pipeline.
We decided to fix it once."

The Arsenal

Four products.
One obsession.

They don't build everything. They build the part everyone else gets wrong.

zerank-2

Flagship reranker. Fastest. Most accurate.

A cross-encoder neural reranker trained with the proprietary zELO method — an Elo-scoring system originally used for chess ratings, now applied to query-document relevance. Beats Cohere rerank 3.5 and Jina rerank m0 in both speed and accuracy. About 12% faster on small payloads, 31% faster on large ones. Available via API, HuggingFace, AWS SageMaker, and Azure Marketplace.

zembed-1

4B params. Tops the leaderboard.

A 4-billion parameter open-weight multilingual embedding model distilled directly from zerank-2 — meaning its relevance intuition is inherited from the reranker itself, not just binary labels. Supports 50+ languages. Compresses from 2,560 dimensions all the way down to 40. Reduces vector storage costs by up to 10x. Outperforms OpenAI, Cohere, Google, and Voyage on finance, healthcare, and legal benchmarks.

zsearch

End-to-end. Ship in an afternoon.

ZeroEntropy's full search engine: ingestion, preprocessing, hybrid retrieval, embedding, and reranking — all in one API. Handles negated queries ("articles NOT mentioning Elon Musk"), multi-hop queries, and fuzzy filtering. A Python SDK or interactive dashboard. No pipeline to stitch. No vector DB to babysit.

ze-onprem

Enterprise. VPC-native. HIPAA-ready.

Run the entire ZeroEntropy stack inside your own cloud. No data leaves your VPC. SOC 2 Type II certified. HIPAA-ready. 99.99% SLA. White-glove onboarding and custom integrations. Available on AWS SageMaker and Azure Marketplace. Private offers for volume pricing and BAAs.

Developer Experience

One import.
Human-level search.

It's embarrassingly simple. That's the point.

PYTHON · 60ms latency
# Create an API Key at https://dashboard.zeroentropy.dev
fromzeroentropyimport ZeroEntropy

zclient = ZeroEntropy()

response = zclient.models.rerank(
model="zerank-2",
query="Which reranker is fastest AND most accurate?",
documents=[
"Cohere rerank-3.5 · 171ms",
"Jina rerank-m0 · 300ms",
"ZeroEntropy zerank-2 · 60ms · #1 NDCG@10",

      ],

    )

# zerank-2 comes out on top. Shocking, we know.

Backed by

Y Combinator

Initialized Capital

a16z Scout

Entrepreneurs First

22 Ventures

Transpose Platform

Angels: OpenAI · HuggingFace · Front

By the Numbers

The benchmark
doesn't lie.

NDCG@10 on MSMARCO — the closest public proxy to real RAG workloads. Higher = better.

⭐ zembed-1 (ZeroEntropy) 0.946

Voyage 4 ~0.93

OpenAI text-embedding-3-large ~0.92

Cohere Embed v3 ~0.89

Gemini text-embedding-004 ~0.87

Source: Agentset Leaderboard & ZeroEntropy public benchmarks. Scores approximate for competitor models.

What You Can Build

Search that earns trust
in industries that demand it.

Legal. Healthcare. Finance. Customer Support. Sales. The ones where wrong answers have consequences.

⚖️

Legal Research

LegalBench-RAG — the first open-source legal retrieval benchmark — was built by ZeroEntropy. 6,800+ queries. 79M+ characters. Human-annotated spans. Contract understanding, case law retrieval, and regulatory search at scale.

🏥

Healthcare AI

Clinical documentation search, diagnostic support, medical literature retrieval. zembed-1 shows particularly strong domain performance on healthcare vocabulary and nuanced relevance ranking. HIPAA-ready on ze-onprem.

📈

Finance & Trading

Earnings call analysis, research note retrieval, risk document search. zembed-1 outperforms every competitor on finance-domain benchmarks. Nicholas has built stat-arb algorithms and audited blockchains — they know this space.

🎧

Customer Support

Assembled — the support platform trusted by Stripe, Canva, Robinhood, and Notion — replaced their entire retrieval stack with ZeroEntropy. Result: better accuracy, lower latency, same scale. Zero rebuilding of pipelines.

🤖

AI Agents

zsearch is purpose-built for agentic workflows. Agents ask weird, multi-hop, negated, fuzzy questions. ZeroEntropy's system routes each query to the right retrieval strategy automatically. No hardcoded logic. Just accurate answers.

🛒

E-Commerce Search

Users scan the top 10 results. Every misranked product is a lost sale. zerank's 60ms latency and class-leading NDCG@10 means better product discovery without the latency tax that kills conversions.

The Journey So Far

A year. A lot of benchmarks.

Small team. Big leaderboard presence.

2024

Company Founded

Ghita and Nicholas, fresh from AI engineering roles across healthcare, finance, and SaaS, decide to fix retrieval once and for all. Pre-seeded by Entrepreneurs First.

Winter 2025

Y Combinator Batch W25

ZeroEntropy joins YC's winter cohort. The combination of math-competition credentials and real production experience stands out in a crowded AI infrastructure field.

Early 2025

zerank-1 & zerank-1-small Launch

First public rerankers. zerank-1-small goes fully open-source under Apache 2.0. zerank-1 benchmarks show NDCG improvements of up to 5 points over Cohere, Voyage, and Salesforce's models. Outperforms Gemini Flash 2.0 as a reranker.

July 2025

$4.2M Seed Round

Led by Initialized Capital. YC, Transpose Platform, 22 Ventures, a16z Scout, and angels from OpenAI, HuggingFace, and Front join. TechCrunch covers the round: "Moroccan founder raises $4.2M for her YC-backed startup."

Late 2025

zerank-2 & Assembled Partnership

zerank-2 ships and immediately tops reranking leaderboards. Assembled — whose platform serves Stripe, Canva, Robinhood, and Notion — switches 100% of production retrieval reranking to ZeroEntropy.

March 2026

zembed-1 — #1 Embedding Model

The 4B open-weight multilingual embedding model ships and immediately tops the Agentset Leaderboard. Outperforms OpenAI, Google, Cohere, and Voyage on general retrieval and multilingual tasks.

Culture & Team

Mathematicians.
Physicists.
Competitive programmers.

ZeroEntropy's HuggingFace page says it plainly: "a team of mathematicians, physicists, and competitive programmers." This isn't branding. zerank-1 was initialized from Qwen3-4B and trained with a novel zELO pipeline derived from the Thurstone statistical model. zembed-1 was distilled from the reranker itself, inheriting calibrated relevance judgments that binary training labels cannot produce.

The hiring bar is explicit: they are actively recruiting a Head of Developer Experience — someone technical who can make complex ideas clear and enjoys being around builders. The model: deep technical, developer-first, no hand-waving.

Ghita uses her platform to speak openly about diversity in deep tech — particularly the near-absence of women in AI infrastructure. Her journey from Morocco to École Polytechnique to YC is part of ZeroEntropy's story, not a footnote to it.

"There aren't many women in DevTools or AI infra. But if you're drawn to complex, technical problems — don't let anyone make you feel like you're not capable." — Ghita Houir Alami, CEO, ZeroEntropy

Open Roles (via YC)

🔧 Head of Developer Experience
⚙️ ML Engineer
🏗️ Backend Engineer
🌐 Go-to-Market / Growth

View Open Roles →

What ZeroEntropy Wants

This page exists for a reason.
Here's what they're after.

🏢 Enterprise Customers

If you are building AI agents, RAG pipelines, internal search, chatbots, or search bars at scale — and your retrieval layer is currently duct-tape and prayers — ZeroEntropy will save you months of engineering. Contact: founders@zeroentropy.dev

🤝 Investors & Partners

The $4.2M seed is deployed. The models are shipping and topping leaderboards. The next round will be oversubscribed. If you want in early on the infrastructure layer every AI product needs — reach out now. AWS, Azure, and HuggingFace are already distribution channels.

👩‍💻 Engineers & Researchers

If you find zELO methodology interesting, if you've published on RAG or retrieval, if you've won a coding competition — they want to hear from you. The team is small. The problems are genuinely hard. The output is open-weight and published.

Fun fact: there's a video game
also called "Not Entropy: Zero".
ZeroEntropy has a whole page
explaining they are not it. 🎮

The Name Game

A physics term. A chess scoring system. A Moroccan founder. A CMU dropout.

The name ZeroEntropy comes from information theory: high entropy = chaos, unpredictability, noise. Zero entropy = perfect knowledge of what a message contains. That's the philosophical claim they're making about search. Not "better search." Certain search. The zELO training method — their most important technical innovation — takes its name from the same Elo rating system used in chess, applied to document relevance scoring. The company is, essentially, a physics-meets-chess-meets-mathematics metaphor made into software.

ZeroEntropy

AI doesn't hallucinate.It retrieves badly.

Two brains.Infinite equations.

Pre-seeded by Entrepreneurs First.Polished by Y Combinator.

Four products.One obsession.

One import.Human-level search.

The benchmarkdoesn't lie.

Search that earns trustin industries that demand it.