Ninatuzi

AI Algorithm Engineer

Tencent Holdings • Full time

May 2026 - Present • 0 mos

Designed and deployed a production-grade Agentic RAG system for an internal lithium battery R&D knowledge platform, serving confidential engineering data securely for internal staff. Architected a six-skill Agent pipeline covering intent recognition, query rewriting, ChatWiki memory retrieval, RAG, answer aggregation, and fallback response. Built a custom ChatWiki memory module that stores each conversation turn as structured {query, answer, retrieved_knowledge} records, combined with LLM-powered topic grouping and two-level vector retrieval (module-level topic grouping → wiki-level semantic search), enabling precise multi-turn context recall without memory contamination from off-topic chitchat. Engineered a hybrid intent recognition system combining domain keyword hard-rules for 60+ battery-specific terms with LLM classification using windowed dialogue history, accurately distinguishing knowledge queries, follow-up questions, and off-topic chitchat — resolving critical edge cases where safety-related queries were misrouted to general knowledge responses. Designed a topic-boundary detection mechanism using LLM-based query rewriting and coreference resolution to eliminate context-drift errors across conversation turns, ensuring the system correctly resets context when users switch topics. Implemented dynamic module truncation in the ChatWiki retrieval skill, retaining only closely-scored topic modules to prevent false context injection while maintaining retrieval efficiency. Constructed an internal knowledge graph over confidential battery manufacturing data to enforce data access control and support structured reasoning within the Agent pipeline. Implemented workspace-level session isolation with automatic cleanup on session deletion to ensure data security for sensitive industrial content. Led full end-to-end delivery including system architecture design, reasoning chain documentation, prompt engineering, multi-dimensional evaluation, and deployment on private infr

AI Algorithm Intern

Alibaba • Full time

Jul 2025 - Nov 2025 • 4 mos

Joined an R&D team exploring how large language models can elevate customer service quality in the travel and tourism industry, contributing across the full pipeline from raw data processing to Agent optimization and system deployment. Designed and executed a comprehensive data governance pipeline to process 50,000+ raw travel domain utterances, applying LLM semantic scoring combined with regex-based filtering to distill 8,000 high-quality instruction pairs for model fine-tuning. Injected local cultural knowledge, religious sensitivity guidelines, and destination-specific expertise into the dataset to improve domain coverage and ensure compliance with local norms. Responsible for Agent decision routing logic optimization, redesigning the intent classification and retrieval chain to improve accuracy and reduce unnecessary LLM inference overhead. Optimized the RAG retrieval pipeline to improve knowledge recall precision, resulting in a 66% reduction in factual error rate compared to the baseline system, and a significant improvement in query automation coverage and user trust. Contributed to the fine-tuning workflow using LoRA on Llama 3.1-8B, supporting data preparation, training configuration, and evaluation. Participated in building a multi-dimensional evaluation framework using GPT-4o automated scoring combined with human validation to measure intent recognition accuracy, factual correctness, hallucination rate, and user satisfaction across different tourism scenarios. Delivered a production-grade intelligent tour guide system with industrial-level reliability, forming the core technical foundation for the company's commercial Agent product launch.