Derived from job-description analysis by Serendipath's career intelligence engine.
Original posting from TikTok
Project Overview
The Global E-commerce ecosystem has accumulated massive heterogeneous data, including user behavior, product images and text, multimedia content, sales data, and logistics time series. Traditional models still face significant limitations in long‑term forecasting, cross‑modal understanding, and complex decision‑making.
This project aims to build a foundational large model tailored for Global E-commerce scenarios. It will unify key elements such as users, products, content, logistics, and inventory into a single modeling framework. On top of this, a modular, pluggable Agent framework will be designed to integrate capabilities such as task planning, tool usage, multi‑turn interaction, and environmental awareness. This enables end‑to‑end intelligent decision‑making across workflows like demand forecasting, traffic allocation, and personalized recommendation.
Key Challenges
- Heterogeneous Data Fusion & Alignment: Unified modeling of user behavior sequences, product sales time‑series signals, and multimodal product content, achieving deep semantic alignment across high‑dimensional temporal and visual/textual representations.
- Collaboration Between Recommendation LLMs and World Models: Reformulating recommendation as a generative problem of producing user‑specific recommendation lists, enabling end‑to‑end modeling based on large models.
- Item Tokenization for Recommendation: Efficiently encoding hundreds of millions of items into multimodal semantic representations to support large‑scale training and generation tasks. Handling tens of terabytes of user behavior tokens during pretraining, improving scaling laws through model architecture and training strategies, reframing recommendation tasks into post‑training problems, and optimizing for GMV and user experience.
- Multimodal Large Models for E‑commerce: Developing multilingual and multimodal large models tailored for e‑commerce, achieving state‑of‑the‑art performance in core scenarios, and serving as the foundation for intelligent e‑commerce agents across diverse applications.
- Agent Evaluation, Safety & Compliance: Designing evaluation metrics and benchmarks aligned with real‑world business scenarios, ensuring robustness, safety, and compliance under highly constrained and adversarial environments.
Project Value
- Technical Value – Build a general‑purpose multimodal foundation model, leveraging iterative improvements in models, data, and compute to achieve scaling‑law‑driven growth and establish a strong technical foundation.
- Business Value – Establish a foundational large model for Global E‑commerce, leveraging generative recommendation, temporal models, and agent‑based systems to drive GMV growth and user retention, forming a high‑leverage revenue engine.
Responsibilities
- Design algorithms and systems that leverage LLMs and generative models for content‑to‑commerce matching, product summarization, etc.
- Explore novel architectures and strategies for generative recommendation systems.
- Contribute to the research community via internal papers, patents, or external publications.
- Drive scientific rigor while balancing real‑world constraints.
Candidate Requirements
Successful candidates must be able to commit to an onboarding date by the end of year 2027. Please state your availability and graduation date clearly in your résumé.
Qualifications
Minimum Qualifications
- Completing or recently completed a PhD in Software Development, Computer Science, Computer Engineering, or a related technical discipline.
- Strong foundation in machine learning, with knowledge of cutting‑edge AI technologies; publications in accredited academic conferences or competition experience are preferred.
- Familiarity with big data frameworks such as Hadoop, MapReduce, and Spark.
- Experience with TensorFlow or PyTorch for model training and deployment; understanding of training acceleration techniques such as mixed precision and distributed training.
Preferred Qualifications
- Knowledge of model compression and inference acceleration techniques, including but not limited to quantization, pruning, distillation, and TensorRT optimization.
- Expertise in at least one of the following areas:
- Computer Vision & Multimodality: In‑depth research experience in multimedia or computer vision fields, including image search, classification, segmentation, object detection, OCR, graph neural networks, multimodal learning, and unsupervised/self‑supervised learning. Experience with large‑scale CV/multimodal models, particularly in e‑commerce scenarios, including developing and optimizing multimodal models for e‑commerce videos and products. Ability to integrate LLMs with video/product representations to support tasks such as multimodal classification, video QA, cross‑modal retrieval, and product categorization with performance exceeding production models. Strong hands‑on experience, with achievements in competitions such as Kaggle, COCO, ImageNet,
Source: TikTok careers