Scaling Fashion AI: How Zhiyi Built a Massive Image‑Recognition Platform on Alibaba Cloud

Company Overview

Hangzhou Zhiyi Technology Co., Ltd is a national high‑tech enterprise driven by artificial intelligence, founded in February 2018. It secured a $10 million Series A round in its founding year and a RMB 200 million Series B round in 2021, earning a place on Hangzhou’s quasi‑unicorn list.

Zhiyi has developed a suite of SaaS products—Zhiyi, Zhikuan, Meinian—providing fashion trend forecasting, design empowerment, and intelligent style recommendation, serving thousands of fashion brands such as UR, Vipshop, and Peacebird.

Solution Architecture on Alibaba Cloud

The platform is divided into product, service, data, and big‑data layers.

Product layer: multiple apps (Zhiyi, Meinian) and customizable APIs for data access and image‑search.

Service layer: front‑ and back‑end systems containerized on Alibaba Cloud Container Service for Kubernetes (ACK).

Data layer: raw images stored in OSS, business data in MySQL, KV data in HBase, feature vectors in Proxima, and metadata in Elasticsearch.

Big‑data components include Log Service (SLS) for vector caching, DataWorks + MaxCompute for offline warehousing, Python recommendation jobs on ACK, and an on‑prem GPU cluster for batch image recognition.

Big‑Data Solution Evolution

Stage 1 – IDC‑built CDH Cluster

Data were synced from Alibaba Cloud to an on‑prem CDH cluster for Hive processing, then results returned to the cloud.

Stage 2 – DataWorks + MaxCompute

Computation migrated to MaxCompute with task orchestration via DataWorks, reducing operational complexity.

Stage 3 – Elasticsearch for Ad‑hoc Queries

Elasticsearch (managed by Alibaba Cloud) with Proxima vector search enables fast multi‑dimensional queries for trend analysis, handling millions of records per query.

Image Recognition Pipeline

All fashion images are processed offline: each image is transformed into a 256‑dimensional feature vector and indexed by Proxima.

Model Training

Annotated images are used to train models on a GPU cluster. Active Learning reduces labeling cost by iteratively training on a small labeled set, predicting unlabeled data, and refining with human verification.

Batch Image Recognition

Trained models are packaged into Docker images and run on GPU nodes. Feature vectors are first cached in SLS, then cleaned (deduplication, etc.) via DataWorks tasks before being written to Proxima. Cloud‑based preemptible GPU instances supplement on‑prem resources, cutting costs.

Single‑Image Recognition

Through an online serving endpoint, users can upload a single image and receive inference results instantly.

Image Search (Visual Search)

When a user uploads a new image, its vector is computed and searched against the Proxima index to retrieve the most similar fashion items, enabling content‑based image retrieval.

Faiss Deployment and Challenges

Initially, Faiss was deployed on a distributed GPU cluster for vector search, but faced stability issues, GPU memory limits, high operational overhead, bandwidth contention, and limited recall size.

Proxima Vector Engine

Proxima, Alibaba’s in‑house vector search engine integrated with managed Elasticsearch, offers high stability, HNSW‑based algorithm without GPU, easy scaling, no bandwidth contention, and larger recall sets, addressing the shortcomings of Faiss.

Technical Roadmap

OLAP Analysis Optimization

Switching from Elasticsearch to ClickHouse for pre‑aggregated wide tables reduced query latency from 9 seconds to under 2 seconds for large‑scale image statistics.

Data Modeling and Governance

DataWorks is being used to enforce unified data modeling standards and automated quality checks, improving data management across the organization.

Future Image‑Search Collaboration

Alibaba Cloud PAI’s similarity‑search solution, which builds models without manual labeling, is under evaluation for potential integration.

More Information

Proxima: https://developer.aliyun.com/article/782391 Image Retrieval Solutions: https://help.aliyun.com/document_detail/313270.html