Technology

Engineered to be measured, observed and reversed.

Feiliu runs on a mainstream, production-grade stack. Every layer is observable and every strategy is versioned.

Architecture

Four building blocks, one feedback loop.

Architecture overview: signals feed online features and vector retrieval, models rank candidates, and experiments feed results back.
  1. Multimodal signals

    Text, image and video understanding fused into shared representations for items and users.

    • Text
    • Image
    • Video
  2. Vector retrieval

    Embedding-based candidate retrieval alongside rule and behaviour-based recall strategies.

    • Embeddings
    • ANN search
  3. Online features

    Streaming feature updates and real-time feedback keep rankings current with user behaviour.

    • Kafka
    • Flink
  4. A/B experimentation

    Traffic splitting, consistent tracking across platforms and comparable results for every strategy.

    • Traffic split
    • Metrics

Stack

Mainstream tools, carefully assembled.

Services
Go / Java
Interfaces
gRPC / HTTP
Streaming
Kafka / Flink
Retrieval
Vector search
Models
PyTorch

Observability and rollback

Know what changed. Undo it if you need to.

  • Latency and SLA monitoring

    P99 latency and service health tracked against the agreed SLA.

  • Strategy audit trail

    Every strategy change is logged with who, what and when.

  • Versioned rollback

    Strategies and models are versioned, so a change can be reverted quickly.

  • Consistent tracking

    The same events and experiment framework on every platform, so results compare.

Bring better recommendations to your platform.

Tell us about your content and audience. We will propose a proof of concept scoped to your goals.