Samuel Herman

I build AI that retrieves, understands, and sees: the science, the models, and the end-to-end capabilities that make it work at scale and in the physical world.

LinkedIn GitHub arXiv

Retrieval science

Predicting how an approximate index will behave before it is built, from the geometry of the embeddings alone. Recall forecasts land within 0.03 of measured on unseen million-document corpora.

Language models

Adapting embedding and sparse models to vocabulary they were never trained on, and pricing multi-vector and late-interaction retrieval at production scale.

Visual intelligence

Computer vision and visual document retrieval: models that read pages and images directly, indexed and served at the same economics as text.

AI in the physical world

Perception systems that see and act at the edge, measured on what they get right per watt and per dollar.

Research: Closed Forms and Synthetic Twins (arXiv:2609.00364, cs.IR, August 2026) predicts approximate-index recall from label-free embedding statistics, within 0.03 of measured recall on an unseen million-document corpus.
Open source: Member of the OpenSearch Leadership Committee. Built lock-free concurrent graph construction for the jVector plugin for OpenSearch.
Industry: Led engineering organizations of more than 75 people at AWS, Oracle Cloud Infrastructure, and DataStax, building search, SIEM, and observability systems that process petabytes a day.
Patents: Patents in distributed systems and storage technologies.

Latest research

The paper's two headline facts: the fidelity/efficiency trade for PQ and FDE corrections, and synthetic-twin recall predictions for HNSW and IVF across 17 encoders — FigureLeft pair: the fidelity/efficiency trade drawn as a before-and-after move; each panel starts at its corpus's native baseline (open square). On the dense corpus the correction is free: the OPQ rotation gains 0.16 efficiency with fidelity kept. On the late-interaction corpus the correction costs fidelity and pays: centering spends 0.08 fidelity for 0.42 efficiency, while whitening (dashed) buys the same gain for 0.15. Right: recall simulated on a cluster-matched synthetic twin, built purely from label-free statistics with no real index, predicts measured recall for both HNSW and IVF across 17 encoders with zero fitted parameters.

Closed Forms and Synthetic Twins: Predicting Approximate Nearest Neighbor Recall from Embedding Statistics

arXiv:2609.00364 [cs.IR] · August 2026 · Shmuel Herman

Embedding models are benchmarked as if retrieval were exact. In production they sit behind HNSW, IVF, product quantization, or fixed-dimensional encodings whose recall the benchmarks never measure. The paper shows that recall is predictable before the index exists, from one cheap measurement pass over the raw embeddings, and that the same geometry is trainable.

arXiv PDF Paper page and BibTeX

Selected writing

All writing, in order

Talks

2026

From Embeddings To Index: A Practitioner's Guide To Domain-Adapted Neural SearchOpenSearchCon, Prague

Watch 2025

Vector Search Beyond The Hype: OpenSearch and Vector DatabasesOpenSearchCon

Watch 2024

Lucene And Beyond: Core Storage Extension In OpenSearchOpenSearchCon

Watch 2023

Journey To 750 Node ClustersOpenSearchCon

Watch

All talks