HIEU NGUYEN

• Led the adoption of AI-assisted coding at the company by proposing and guiding the implementation of Claude Code within the engineering workflow.
• Applied specification-driven development and test-driven development practices to accelerate implementation while keeping review, validation, and delivery quality under control.
• Built LLM-powered Q&A agents for public service platforms used by provincial administrative agencies and the Ministry of Public Security of Vietnam, improving accessibility through natural language interactions.
• Hosted and served LLM workloads with vLLM and SGLang to improve inference throughput and simplify production deployment.
• Architected a multi-agent workflow where a supervising agent coordinates domain-specific agents to handle complex requests with clearer delegation and maintainability.
• Designed a knowledge-enhanced QA system backed by an automatically constructed knowledge graph to support explainable search over political figures and related information.
• Deployed the full stack on-premise with local LLMs to meet data security and privacy requirements while avoiding dependency on third-party APIs.

• Trained Dama 2 7B from scratch for Vietnamese on the Llama 2 architecture, placing 2nd on the VLSP 2023 LLM benchmark.
• Developed Phi-3 Vietnamese and Mistral 7B Vietnamese variants to improve math reasoning, code generation, multi-tasking, and structured outputs for Vietnamese users.
• Fine-tuned and evaluated large language models across multiple tasks, including instruction following, function calling, and JSON generation.
• Worked with high-performance GPU infrastructure to train and iterate on large-scale language models efficiently.

• Developed a TrOCR-based OCR system for English and Vietnamese text, improving accuracy over Tesseract and ABBYY OCR.
• Built a Transformer-based Vietnamese spell-checking model with strong detection and correction performance on the VSEC benchmark.
• Applied few-shot prompting and extraction guidance to convert unstructured documents into structured data formats.

AI researcher & AI Engineer

• Technologies: Python, LangGraph, Neo4j, Amazon S3 Vectors, FastAPI, Next.js
• Team size: 1
• Designed and built a legal knowledge graph that links articles, amendments, references, and regulatory documents for explainable retrieval.
• Used Amazon S3 Vectors as the vector storage layer to support semantic retrieval and multi-step reasoning over legal documents.
• Combined symbolic retrieval with LLM-based reasoning to improve answer accuracy and transparency for legal question answering.

Freelance - AI Developer & AI Engineer

• Technologies: Python, PyTorch, Transformers, FastAPI
• Team size: 3
• Fine-tuned machine translation and summarization models for the technology domain in English and Vietnamese.
• Built the API layer for model serving with load balancing and dynamic batching to support reliable inference.

AI Developer & AI Engineer

• Technologies: Python, PyTorch, Transformers, Vision Transformers, LangChain, OpenCV, FastAPI, Django, MySQL, MinIO, Redis, Elasticsearch, Prometheus, Grafana
• Team size: 6
• Trained an OCR model from scratch with the TrOCR architecture, improving CER by 2% and WER by 9% over Tesseract and ABBYY.
• Deployed the model through a scalable API with load balancing, dynamic batching, caching, and 8-bit quantization, reducing inference time by 50% and memory usage by 4x while preserving 98% accuracy.
• Applied few-shot prompting and extraction guidance to convert documents into structured outputs, improving accuracy by 10% over prior transformer-based approaches.

AI Developer & AI Engineer

• Technologies: Python, PyTorch, Transformers, FastAPI, Django, MySQL, MinIO, Redis, Prometheus, Grafana
• Team size: 6
• Developed a Transformer-based Vietnamese error-correction model with strong detection and correction scores on the VSEC benchmark.
• Deployed the model through a scalable API with load balancing, dynamic batching, caching, and post-training optimization.

• Deploy a language resource, and use prebuilt models
• Create a custom text classification solution
• Create a custom named entity recognition (NER) solution

• TOEIC 640

• Awarded scholarships for the academic years 2019-2020 and 2020-2021