LEQUAN WANG

About

LLM/Conversational AI engineer who built a production multi-turn dialogue system using LoRA fine-tuning on LLaMA, cutting training compute/memory by 90%+ while maintaining low-latency inference via quantization and streaming generation. Experienced in orchestrating end-to-end ML workflows with Prefect/Airflow/Kubeflow (including hyperparameter sweeps and W&B tracking) and improving agent reliability through benchmark-driven testing, shadow-mode rollouts, and stakeholder-informed guardrails.

Experience

Education

Key Strengths

• Built and deployed LoRA-fine-tuned LLaMA conversational system reducing training compute/memory by >90%
• Designs end-to-end LLM pipelines (data prep, training, evaluation) with Hugging Face and PEFT
• Inference optimization on constrained hardware (quantization/bitsandbytes, batching, streaming generation)
• Robust handling of noisy multi-turn conversational data via automated cleaning/normalization and consistent prompt formatting
• Reliability engineering for AI agents: guardrails, unit/integration tests, benchmark suites, shadow-mode + human-in-the-loop rollout
• Strong model/method selection using empirical benchmarks (task success, hallucination rate, latency) and cost/performance trade-offs
• Workflow orchestration for ML (Prefect/Airflow/Kubeflow) including scheduled pipelines, sweeps, and metric-driven checkpoint selection
• Effective collaboration with non-technical stakeholders by translating feedback into measurable evaluation criteria and scenarios