He Tianqi

Owned end-to-end RL and LLM post-training pipelines for Huawei Ascend ecosystem, including data curation, training, and evaluation. Built agentic RL post-training for Qwen2.5-coder 7B with vLLM/Ray and 16K context with 16-turn tool-calling, stabilizing distributed training and improving AIME24 while avoiding degradation on AIME25. Deployed GRPO-based RL for Qwen3-30B-A3B MoE on a 48-NPU Ascend cluster with decoupled resource allocation and concurrency optimization. Implemented multimodal RLHF (T2V) for HunyuanVideo-13B with dynamic CPU/NPU offloading to resolve OOM constraints, improving iteration time and enforcing alignment checks using VBench. Executed 8-bit/4-bit low-bit pre-training for OLMo 1B on 300B tokens, refactoring for Online Softmax and validating via lm_eval across perplexity and multiple zero-shot benchmarks.

Built multi-agent workflows using LangChain and LangGraph for investment Q&A and natural-language trading. Designed stateful agent routing that decomposes user intents into executable DAGs of tool calls, enabling robust multi-step reasoning and dynamic tool execution. Engineered a streaming pipeline with Kafka and FastAPI to automate social media copywriting, replacing manual prompting with dynamic templates backed by internal APIs. Created an evaluation and observability platform with LangSmith for A/B testing and rubric-based Judge LLM evaluation, achieving high regression test coverage. Orchestrated BI reporting workflows with Airflow, automating scheduled Feishu group reports and saving significant monthly effort.

Developed data production workflow for ChatGLM coding-focused training, including technical guidelines and validation scripts to ensure correctness for code generation and bug-fixing tasks. Established standardized quality criteria and performed sampling-based inspections to filter low-quality annotation data using logical consistency and functional correctness checks.