Renyuan LIU (刘稔rěn远)

Project Background: As a core developer, I worked on an e-commerce shopping Agent spanning 100 tools and eight business domains, addressing full-tool overhead, confusion among similar tools, incorrect ordering of multi-intent requests, and cross-turn history contamination.

• Intent Routing (100.0% Classification Accuracy): Built 48K e-commerce and casual-conversation samples and LoRA-tuned Qwen3-0.6B, raising accuracy on 9,000 isolated test cases from 84.8% to 100.0%. Compared with Qwen3-1.7B, median inference latency fell by 34.3% and throughput rose by 52.0%.
• Efficient Tool Retrieval (97.4% Top-10 Full Tool Recall): Built 70K positive pairs and 96K typed negatives, fine-tuned bge-large with multi-negative contrastive learning, and trained bge-reranker-v2-m3 with 1:7 positive-to-hard-negative groups. Implemented dual BM25/vector recall, two-stage RRF fusion, and Cross-Encoder reranking with a “recall 40 → rerank 20 → plan 10” funnel. Top-10 full recall rose from 47.6% to 97.4%; compared with “rerank 30 → plan 20,” planner input tokens and median reranking latency fell by 36.8% and 33.7%.
• Multi-Step Tool Planning (92.6% End-to-End Accuracy): Built 78K multi-style tool-call trajectories and LoRA-tuned Qwen3-8B with MS-SWIFT to output ordered tool chains from a fixed Top-10 candidate set. Constructed 4,000 preference pairs from real parent-model errors, using preference optimization followed by low-learning-rate DPO for convergence. Exact tool-chain match improved from 85.7% to 88.9%, with 92.6% end-to-end accuracy on 9,000 static tests.
• Multi-Turn Agent Framework (92.8% Context-Dependent Request Accuracy): Designed an external context orchestrator whose rule-based state machine schedules recent state, rule summaries, and in-session BM25 retrieval. It compresses necessary history into a standalone current request before invoking the original planner, while request boundaries and full-turn recovery suppress history contamination. A 12-stage ablation on nearly 2,000 natural conversations of 2–7 turns raised independent-test overall accuracy from a 50.5% direct-concatenation baseline to 94.6%; context-dependent and context-independent request accuracy reached 92.8% and 97.4%, while version-validation caching reduced planner calls by 60.3%.

Core Results: 100.0% intent accuracy on 9,000 isolated tests; 97.4% Top-10 full tool recall (47.6% baseline); 92.6% end-to-end accuracy; 92.8% context-dependent request accuracy; and 60.3% fewer planner calls.

Project Background: As a core developer, I participated in the full-parameter post-training of DeepSeek-V4-Pro on domestic accelerators, jointly conducted by Shenzhen Loop Area Institute, Huawei, and 智子芯元. For DeepSeek-V4 inference and post-training on domestic hardware, I worked on Ascend 910C deployment evaluation, kernel-optimization Agent development, and fused backward-kernel optimization.

• 910C Inference Adaptation and Evaluation (DeepSeek V4 Flash/Pro): Participated in deploying and evaluating DeepSeek V4 Flash/Pro on an Ascend 910C cluster. Compared Prefill/Decode performance between vLLM Ascend and SGLang on ShareGPT and other datasets to support inference-framework selection and deployment configuration.
• Kernel Optimization Agent (Ascend Migration and Open Source): Addressed the manual analysis and repeated trial-and-error required by Ascend kernels by rebuilding MIT Kernel Design Agents for 910B/910C with a kernel knowledge base, optimization rules, and performance-diagnostic tools. Connected retrieval → generation → compilation → accuracy/performance verification → retain/rollback into a closed loop and open-sourced Kernel-Agent-Ascend.
• Fused Backward-Kernel Optimization (20.9% Lower Latency): Used msprof to locate scalar address calculation on AIV, sparse memory access, synchronization waits, and load imbalance. Simplified address calculation, pruned low-hit data-movement branches, removed redundant synchronization, optimized tiling/on-chip caching, redistributed vector tasks, and fixed tail-block out-of-bounds access introduced by cache expansion. Kernel latency fell from 22.603 ms to 17.88 ms (1.26×), while cosine similarity remained 1.0 for all four outputs.

Core Results: Under the target configuration (BSND/BF16, B=1, S=4096, K=1024), fused backward-kernel latency fell from 22.603 ms to 17.88 ms, a 20.9% reduction (1.26×). All four output cosine similarities were 1.0, with a maximum relative error of 8.39e-4.