Akulaku’s Adoption of OpenMLDB: Business Scenarios, Technical Architecture, and Evolution Recommendations

Akulaku, a Southeast‑Asian fintech founded in 2016, handles massive transaction volumes across credit‑card‑like services, virtual credit cards, banking, and investment products, requiring robust risk‑control and AI technologies such as computer vision, NLP, and graph analysis.

The company’s machine‑learning stack is organized into a scenario layer (CV, OCR, speech, graph mining, AutoML) and a platform layer where data processing consumes over 80% of resources, while service and algorithm layers use standard industry solutions.

Akulaku first learned about OpenMLDB in early 2021, evaluated its source code, and after several months began deploying it in production for both offline feature development and real‑time serving, replacing previous solutions like Neo4j and custom Spark pipelines.

OpenMLDB provides a unified feature‑development system, a consistent technology stack for both offline and online workloads, and guarantees logical consistency and comparable performance between batch and streaming execution, eliminating the need for separate tools and reducing engineering overhead.

Key business scenarios demonstrated include:

RTP sorting service: OpenMLDB processes incoming Kafka streams, generates features, and performs Top‑K ranking with sub‑millisecond latency, outperforming MPP and Flink solutions.

Complex spatial calculations: OpenMLDB’s LLVM‑based just‑in‑time compilation achieves ~30 ms query time on billions of GPS records, over 30× faster than traditional databases and 10× faster than Spark.

Gang‑mining fraud detection: Real‑time graph‑based feature computation is achieved within tens to a few hundred milliseconds, meeting hard‑real‑time requirements that previous offline or Flink‑CDC approaches could not satisfy.

Cost analysis shows an annual infrastructure expense of roughly ¥2 million, while savings from reduced hardware and labor exceed ¥4 million, highlighting significant ROI.

Evolution suggestions from model‑development, data‑development, and operations perspectives include expanding SQL support for multi‑table joins, integrating statistical/model functions as SQL calls, simplifying cold‑start data preparation, extending time‑window lengths, offering cloud‑native deployments, and adding richer auditing, logging, and permission management features.

The content is based on a talk by Akulaku’s Algorithm Director at OpenMLDB Meetup No. 1 and aims to share practical insights for the fintech and data‑engineering communities.