Phạm Trường

At XBrain, I designed an e-commerce shopping copilot integrating Amazon Nova LLMs, session memory, tool calling, and layered guardrails. I connected it to internal microservices via gRPC and combined semantic search with SQL queries for inventory and order retrieval.

I also developed a validation layer for LLM outputs and an LLM-as-a-Judge pipeline using Claude 3.5 Sonnet on AWS Bedrock. The guardrail layer reduced Attack Success Rate to 0%, with a 0.35% False Positive Rate.

On my Course Recommendation Agent project, I built an autonomous recommendation system with LangGraph, Gemini, Qdrant, and BGE-M3. It achieved a 74.3% Hit Rate@5 and an NDCG@5 of 0.67, with 100% output schema compliance and input-stage defense against prompt injection and PII leaks.

I also fine-tuned Wav2Vec2 for British English phoneme-level pronunciation assessment, using SpeechOcean768 for training and L2-ARCTIC for evaluation. The model reached a Pearson correlation of 0.73 and a Spearman correlation of 0.69 against expert human raters.