Henry Bùi

I build LLM agents that hold up in production — grounded in real source code through parsers, retrieval and MCP tools, with guardrails that make wrong answers structurally impossible rather than merely discouraged. 1.5+ years at FPT Software across agent orchestration (LangGraph, AutoGen), RAG and retrieval, tool calling via MCP, agent memory and LLM evaluation; backend in Python and FastAPI with Azure OpenAI, Redis and PostgreSQL.

Measured results: hallucinations −60%, token cost −20–30%, unattended agent completion ~80%.

AI Engineer (Intern Software Engineer, Feb – Aug 2025) at FPT Software (2025-02 – Present)

• Raised first-try success from ~50% to ~80% for a 9-step multi-agent workflow that writes software documents (requirements → design → mockup) by redesigning the workflow.
• Built 6 SDLC agents — brainstorming, requirements, architecture, UX, mockup, PRD review — with personas, skills and reference material.
• Added per-step model routing so each agent uses the model that fits it, balancing cost and quality.
• Lifted unattended agent-job completion to ~80% with timeouts for long runs, recovery after disconnects and no duplicate runs.
• Reduced developer effort by ~85% (5–8 manual steps → 1) by integrating AI agents into Azure DevOps — developers assign tasks by @mention and get results on the work item.
• Kept all project data inside the customer's own network by adding self-hosted Azure DevOps Server support alongside cloud.
• Redacted PAT tokens from logs and traces across both cloud and on-prem connections.

AI Engineer (Intern Software Engineer, Feb – Aug 2025) at FPT Software (2025-02 – Present)

Project: Metis – CodeIQ

• Lowered token cost per code lookup by 20–30% after benchmarking three retrieval methods (regex, code graph, embedding search) on 15–20 repos, one with 3M lines.
• Caught a quality issue before release: embedding search missed real program flow on bug fixes. Proposed hybrid retrieval instead.
• Built an eval set for code search so every change is measured, not guessed.

AI Engineer (Intern Software Engineer, Feb – Aug 2025) at FPT Software (2025-11 – 2026-04)

Project: Legacy PGM – xMainframe

• Parsed 100% of files across ~3M lines with ANTLR4 parsers that turn COBOL/JCL/CopyBook source into structured JSON — programs, variables, dependencies — the grounding data the LLM works from.
• Reduced hallucinations by ~60% (unsupported answers ~50% → ~20% on a 10–20-question AWS CardDemo benchmark) by developing 7 of the platform's 20+ MCP tools on that data — dependency extraction, CopyBook variable length, copybook/COBOL/JCL relationships — served to the LLM via tool calling.
• Reduced repeat questions by ~75% by building a custom agent memory layer — knowledge graph for exact dependencies, vector retrieval for semantic lookup, source fetched last — replacing an off-the-shelf library.
• Reached 70–80% accuracy for AI-generated program flow diagrams against real code; served 5–6 engineers.

AI Engineer (Intern Software Engineer, Feb – Aug 2025) at FPT Software (2025-02 – 2025-08)

Project: Mainframe Studio

• Parsed mainframe languages with ANTLR4 in Python (and Java for performance on larger codebases), extracting program metadata — variables, structures and dependencies — into structured formats for reporting and system integration; built working knowledge of grammar rules, abstract syntax trees and compiler design.
• Applied the AutoGen framework to build multi-agent systems in Python capable of task coordination, information retrieval and basic decision-making, integrated with external tools and APIs.

Information Technology Engineer in Information Technology – Ho Chi Minh City University of Natural Resources and Environment (2020 – 2025)