Hieu Trung Dao

• AI-Driven Engineering Practice: Transformed the engineering and research workflow by embedding AI agents across the full development lifecycle. Automated literature and technology research to accelerate decision-making; used agentic planning for task decomposition and sprint structuring; applied AI-assisted coding, test generation, and code review in daily practice. Deployed real-time monitoring agents that detect, diagnose, and respond to application issues as they occur, and integrated AI into CI/CD pipelines to enable continuously validated, self-improving delivery.
• Legacy Modernization / XMainframe (2023–Present): Led development of an agentic AI platform for mainframe maintenance and modernization (COBOL/JCL/PL/I) spanning the maintenance SDLC: code comprehension and documentation, change impact analysis across programs, copybooks, and JCL, bug investigation via data-flow tracing, refactoring planning, and test-gap analysis. Engineered long-horizon agentic execution over large legacy codebases — context management, tool-result budgeting, middleware hooks, prompt caching, and reliability guardrails — with multi-agent orchestration over custom static-analysis tools and real-time streaming interfaces. Built environments for agent trajectory generation and data-centric optimization: an emulated mainframe runtime (COBOL compilation and execution with CICS, JCL/JES, datasets, Db2, IMS, and MQ) for running legacy workloads off-host, and a behavior-verified modernization pipeline (analyze → capture/replay → rewrite → verify → cut over) built on differential equivalence testing.
• XMainframe LLM Specialization: Adapted LLMs to the mainframe domain, from the original XMainframe model and MainframeBench to successive generations of open-weight models. Built a multi-stage alignment pipeline: LoRA SFT, MoE-aware SLERP merging and model soups to recover base capabilities lost to catastrophic forgetting during fine-tuning, ORPO preference optimization, and GRPO-family reinforcement learning (DAPO) with verifiable code-execution rewards. Established an evaluation methodology pairing academic benchmarks with conversation replay on real production traffic, which exposed benchmark–product divergence and made "beat the untouched base" a gate for every training stage — the result behind the text-space direction above.
• AgentVista / Metis (2025): Architected an enterprise agentic AI platform with state-machine-based workflow orchestration and hybrid RAG pipelines combining vector search and knowledge graphs for code review, documentation, and knowledge extraction.
• SoundAI & SenseAI — Time Series MLOps (2022–2023): Delivered a serverless MLOps platform for acoustic anomaly detection in manufacturing, including stream and batch pipelines, model operations, and automated deployment across edge and cloud environments.
• TBA Xray i2 — Computer Vision (2022–2023): Built industrial computer vision for X-ray quality inspection in FPT's first project with LandingAI. Developed object detection models for defect localization and a multi-object tracking pipeline maintaining consistent item identities across frames. Implemented rule-based inspection logic and delivered an edge inference pipeline on NVIDIA Jetson hardware at production throughput.
• Fine-tuned domain-adapted LLMs on curated datasets; optimized inference serving for production-scale throughput and latency requirements.
• Built multi-stage data and model pipelines; established experiment tracking and model registry workflows adopted across AI Center projects.
• Fine-tuned domain-adapted LLMs on curated datasets; optimized model serving infrastructure for production-scale throughput and latency.

• AI20K is a competency-based AI training program at VinGroup designed to develop practical skills in building and deploying AI agents in enterprise environments.
• Teach and mentor practitioners across two competency tracks: AI Infrastructure & Data (data pipelines, MLOps/LLMOps, monitoring) and AI Application (LLM engineering, agentic design, RAG, multi-agent systems, computer vision, evaluation).
• Teach practitioners across two tracks — AI Infrastructure & Data and AI Application — covering LLM engineering, agentic design, multi-agent system deployment, and computer vision for VinGroup's enterprise AI program.
• Lead lab sessions combining hands-on exercises, project builds, and mentor reviews across the Foundation, Specialization, and Enterprise Practicum phases.

• Competitive research residency focusing on applied AI and production ML systems
• Trained and mentored by PhD-level researchers and global AI experts

Anh T. V. Dau, Hieu Trung Dao, Anh Tuan Nguyen, Hieu Trung Tran, Phong X. Nguyen, Nghi D. Q. Bui

• Developed a domain-specialized LLM for COBOL/JCL comprehension: 77.89% MCQ accuracy (+30% over DeepSeek-Coder) and 62.58 BLEU-4 on code summarization (~6× GPT-3.5)
• Created MainframeBench: first comprehensive benchmark for mainframe code understanding across COBOL, JCL, and PL/I
• Global media coverage: MarkTechPost (US), ITmedia AI+, Ledge.ai (Japan)