Senior / Staff SLM & VLM Engineer - Post-Training, Tool Calling & Agents

Job Summary We are looking for a highly capable engineer/researcher to lead the R&D of Small Language Models (SLMs) and Vision-Language Models (VLMs) for edge / low-latency and cost-efficient production scenarios. You will own the continuous pretraining, supervised instruction tuning (SFT), and compression/distillation pipelines, and work closely with platform teams to deliver reliable, measurable improvements in inference efficiency, tool-use success rate, and overall model quality. Key Responsibilities 1) SLM/VLM Training: Continuous Pretraining & Instruction Tuning (SFT)

• Conduct continuous pretraining and SFT for SLMs and VLMs to improve task performance and domain adaptation.
• Build reproducible training workflows in PyTorch, including data processing, training, evaluation, and model versioning.

2) Compression, Distillation & Edge/Low-Latency Inference Optimization

• Design and implement efficient compression strategies for SLM/VLM, including knowledge distillation, pruning, and quantization-oriented training or post-training optimization.
• Optimize model serving and inference for low-latency / edge scenarios by improving throughput and cost-per-token via techniques such as quantization, caching/KV optimizations, batching strategies, and decoding-time optimizations.

3) Tool Calling System: Catalog, Routing, Validation, Fallback & Observability

• Architect and implement a production-grade tool calling (function/tool calling) framework:
• Tool cataloging and metadata/schema design
• Tool selection/routing and argument construction
• Parameter validation, result verification, and safe fallback/retry strategies
• Call-chain tracing, monitoring, and observability to improve success rate and ROI

4) RL & Reward Modeling for Alignment and Tool-Use Reliability

• Apply post-training methods such as PPO / DPO / GRPO-like optimization and reward modeling to align the model toward objectives including:
• semantic understanding
• tool-use success rate
• content generation quality and consistency
• Support both offline and online iteration loops, including policy evaluation, regression checks, and safe deployment gating.

5) Data Pipeline Automation (Collection, Cleaning, Curation)

• Design automated pipelines for data collection, filtering, cleaning, de-duplication, labeling/weak supervision, and dataset version management to continuously improve training quality.
• Ensure datasets support both SFT and preference/RL style post-training.

6) Rigorous Evaluation, Testing & Iteration

• Build robust evaluation mechanisms: offline benchmarks, task suites for tool-use, regression tests, and reliability metrics.
• Drive rapid iteration through A/B comparisons, ablations, and failure analysis, improving both quality and efficiency over time.

Required Qualifications

• Strong software engineering skills in Python and C++, including experience building ML training/evaluation pipelines in PyTorch.
• Hands-on experience in model efficiency and inference optimization (e.g., distillation, quantization, pruning, serving optimization).
• Experience with high-performance computing and acceleration: CUDA and/or SIMD, profiling and performance tuning.
• Ability to read and reproduce key ideas from recent papers and implement algorithms with strong experimental discipline.
• Ability to communicate effectively in both Chinese (Mandarin) and English as the successful person will have to liaise with counterparts in China.