Job description
---
The team is building a generative recommendation system aimed at financial scenarios. The tech stack closely aligns with the industry SOTA for 2025-2026, emphasizing the deep integration of causal signals with the recommendation pipeline, rather than traditional feature concatenation modeling. You will be directly involved in the design and implementation of the core engine, rather than peripheral data labeling or parameter tuning tasks.
Your work will involve:
Generative recommendation backbone: Implementing user behavior sequence modeling based on HSTU (Hierarchical Sequential Transduction Unit), running through the complete link from recall to ranking; understanding core mechanisms such as pointwise aggregated attention, U-gating, and relative attention bias, and being able to iterate on them.
Semantic IDs system: Using RQ-VAE for residual quantization coding of targets, generating hierarchical semantic IDs; exploring non-general practices of incorporating structured topological features (graph neighborhoods, degrees, PageRank, etc.) into coding inputs.
Causal signal embedding: Truly embedding causal inference into the recommendation pipeline—not just decorating at the explanation layer, but using graph edge weights as hard inputs for recall structure and ranking features, and employing methods like IPS/DR and Uplift CATE for offline policy evaluation and optimization.
Features and online services: Participating in feature engineering and online feature services, collaborating with real-time data infrastructure (stream processing, vector retrieval, graph databases) to support high-frequency recommendation scenarios.
We hope you have:
A solid foundation in machine learning and deep learning, familiar with classic and cutting-edge methods in recommendation systems (sequential recommendation, generative retrieval, recall/ranking paradigms).
Proficiency in using PyTorch, capable of reading and reproducing paper-level implementations (such as Google TIGER and other open-source projects), able to independently complete the full process of data processing, training, and experimental iteration.
Strong self-motivation, able to independently drive tasks in a small team, fast-paced, zero-to-one building environment.
Preference for those with knowledge or interest in causal inference, knowledge graphs, or quantitative finance in any direction.