LLM Security Researcher | AI for Systems Engineer | Independent Developer 🎯 Currently seeking LLM Algorithm/Research positions in Shanghai/Hangzhou

• 🔭 Working on: LLM Security (Jailbreak/Red Teaming), RAG Optimization, LLM Post-training & Alignment.
• 🌱 Studying: LLM Post-training, RLHF/DPO, AI for Systems.
• 💬 Ask me about: LLM Jailbreak, RAG, PEFT (LoRA/QLoRA), Bayesian Optimization.
• 📫 Reach me at: edith_zhang@outlook.com

Competition 1st Place | Developed Hybrid RAG attack framework combining 15 LLMs with QLoRA. Proposed PAP + Safe2Harm strategy, improving safety score by +75% and overall baseline by 916.9%.

📉 PPL 254.40 → 6.06 | Systematic comparison of Zero-Shot, Few-Shot, and QLoRA on Qwen3-4B and Llama-3.1-8B. Established optimal hyperparameter configuration through 17 controlled experiments.

🎯 Recall@10 +4.6% | MRR@10 +6.9% | Built end-to-end RAG pipeline with Bi-Encoder/Cross-Encoder fine-tuning and hard negative mining. Implemented RL-based dynamic Top-M selection, reducing context cost by 66.67%.

️ ASR + LLM + TTS Pipeline | Full-stack independent development (160K+ lines). Integrated Whisper, DeepSeek-R1-Distill-Qwen-14B, and GPT-SoVITS to build a complete voice interaction MVP.

• Jailbreak Olympics:Building & Breaking Safety Systems
• Classical Chinese Instruction Tuning
• RAG System Model Training
• 李宏毅-ML2022-HW15-Meta Learning
• 李宏毅-ML2022-HW14-Lifelong Learning