Fashion AI Infrastructure

The intelligence layer
fashion commerce runs on.

Search that converts. Catalogs that enrich themselves. Trends called before they peak, city by city.

Book a call Get in touch

Live with paying customers

The Label LifeOutzidrOziChumbak

In deployment

A Fortune 500 marketplace A top-10 US fashion retailer
>

Proof, before the pitch

Same catalog. Same query.
Different company.

Generic search matches the words and returns black dresses. Fashion-native search understands that neon accents is the point of the query.

"black dress with neon accents"

Generic keyword search: black dresses, none with neon accents
A leading fashion search engineBefore
Hopit search: black dresses, each with a visible neon element
Hopit agentic searchWith Hopit

Catalogs that enrich themselves

Typical catalogWith Hopit
Attributes per SKU20–30100+
Attribute accuracy~85%90%+
Taxonomyflat list6-level hierarchy

Modelled business impact from richer product pages: +20–40% AOV, +60–85% zero-result recovery, 2–5× merchandiser throughput. These are projected ranges, not measured results from a named account — unlike the benchmark figures below, which are reproducible.

Mumbai is not Milan

Mumbai · IN
Instagram, 30d+62%
Search intent+48%
Global baseline+22%

One global trend graph tells a Mumbai buyer what Milan wants. We score trends per city and per cohort, so the assortment call is local.

One substrate, eight surfaces

And eight more surfaces,
on the same vectors.

Each surface reuses embeddings you have already paid to compute. Add a capability, not a vendor.

#1open model, LookBench image search
9 / 10benchmarks: rank 1 or 2, full corpus
4 msquery latency, CPU-servable
100+attributes extracted per SKU

See how each surface works in depth →

Searchtext → product, image → productrank 1–2 on five of six benchmarks
Agentic Searchmulti-step, outfit and intent awareanswers, not result pages
Trend Intelligencehyperlocal, city and cohort levelmonths of early warning
TrendsFlow™trend lineage and lifecycleproprietary
Catalog Enrichmentattribute extraction at scale100+ attributes / SKU
Personalizationtaste vectors per shopper+20–40% AOV
Size & Fitfit intelligence, returns riskattack the 30% return rate
Commercial Intelligenceassortment and demandgap analysis in plain English
Visual Similarityphoto → product, dedupe, street-to-shop#1 open model on LookBench

The MODA family

Models you can check.

Every number is measured at full corpus, through one harness, with competitors run under identical protocol — and the benchmarks we lose published alongside the ones we win.

Text to image

MODA Pro Lite+213M
Trained encoder with its serving recipe. Leads the ≤250M class on catalog search — KAGL +10.9%, Polyvore +8.7% over MODA, both significant.
Open weights · open recipe ↗
MODA Duo2 indexes
Routes each query to the constituent that suits its shape. Beats either alone on a mixed workload, at one encoder per query.
Open recipe · open weights ↗
MODA203M
Zero-new-parameter recipe over frozen FashionSigLIP. Strongest of the open models on long descriptions and exact-item search.
Open source · open weights ↗
MODA Prohosted
Rank 1 or 2 on five of six full-corpus benchmarks, at single-model query latency.
Closed · Hopit API

Image to image

Find the product from a photograph: street-to-shop, visual similarity, catalogue de-duplication. Our most downloaded models.

MODA-SigLIP-Distilled203M · 768d
Fine R@1 67.63 on LookBench. Above a closed commercial system and a 1.24B model, at 203M.
Open weights ↗
MODA-Matryoshka203M · 64–768d
Pick your dimension at query time. 256d matches 768d on Fine R@1, so a 3× smaller index costs nothing measurable.
Open weights ↗
MODA-Vision-FP1693M vision · 186 MB
Vision tower only, half precision. For edge and mobile deployment.
Open weights ↗

See the full benchmark tables →  ·  Reproduce them yourself →

Talk to us

Start with a small yes.

Fixed fee, two weeks, your catalog and your metrics. You keep the evaluation report either way.

Prefer to talk first? Book a call — we bring specific examples from your category.

Your catalog never trains our public models. VPC deployment available. Deletion on exit, in writing.

We reply personally, usually within a day.