duongnghia222@agentity:~$ whoami

Duong Duc Nghia

AI Engineer

Multi-agent systems & AI-serving infrastructure

Available for freelance & contract work

duongnghia222@agentity — terminal

Type ls to explore, tap a chip, or scroll ↓

// about

AI Engineer with 2+ years building multi-agent systems and AI-serving infrastructure in Python, Go, and Rust. I deliver end-to-end systems — agent runtimes, LLM backends, and RAG pipelines — across fintech, conversational AI, and developer tooling.

Ho Chi Minh City, Vietnam

// experience

AI Engineer · PT Ringkas Asia Technology

current

Mar 2025 – Present · Ho Chi Minh City, Vietnam

›Built a production multi-tenant agentic AI platform for mortgage operations — agent framework, orchestration service, MCP tool servers, and an agent-memory layer — configurable by non-engineers through a back-office.
›Designed a backend-agnostic agent framework: YAML-defined agents with swappable backends (LangGraph, Claude Agent SDK, Google ADK), MCP tool integration, and multi-agent composition with delegation.
›Built the agent orchestration service — a FastAPI gateway over ARQ async workers with SSE fan-out via Redis Streams, multi-tenant run lifecycle (heartbeat + reaper), and AG-UI streaming.
›Built an MCP server exposing 50+ mortgage-domain tools behind two-layer JWT (RS256) + end-user-token auth with response secret-sanitization.
›Built a context/memory service: typed agent memory with hybrid BM25 + pgvector recall fused via Reciprocal Rank Fusion, multi-tenant session state, and conversation compaction.

PythonLangGraphMCPFastAPIRedispgvectorGCP

AI Engineer (Part-time) · Morcalab

current

Nov 2025 – Present · Remote

›Built a solo cross-platform AI desktop app (100K+ LOC, Tauri 2, Rust + React 19/TypeScript): a Rust host orchestrating five sidecar processes over stdio with newline-delimited JSON.
›Architected a multi-agent runtime — two independent kernels (interactive + autonomous) over the AG-UI protocol with tool-approval gating, self-learning memory/skill stores, and a scheduled curator.
›Built a centralized in-host LLM broker: per-feature API-key registry, budget enforcement on a SQLite ledger, and a streaming proxy across seven providers (OpenAI, Anthropic, Google, MiniMax, Ollama).
›Authored LSD, a Go gRPC backend (eight services) re-implementing a LangGraph-style design — assistants, threads, runs, crons, checkpointer, cache — reaching API parity with the Python core on Postgres (pgvector) + Redis.
›Authored code-distilling, a Claude Code / Codex plugin codifying a disciplined five-stage workflow for porting features from reference repos with human review gates.

RustTauriReactGogRPCLanceDBSQLite

AI Engineer · Lumine AI

Mar 2024 – Mar 2025 · Ho Chi Minh City, Vietnam

›Designed and built a scalable AI companion chat platform integrating multiple LLMs for emotionally responsive 1v1 and group conversations, with AI image-generation workflows.
›Implemented long-term memory with Redis and ChromaDB for personalized, context-aware conversations across sessions.
›Built content-moderation and safety filtering with context-aware classification and personalized response generation.
›Researched and integrated TTS and voice cloning to enable spoken interaction.

PythonLLMsRedisChromaDBTTS

// projects

MEP Drawing Analysis System

Freelance

Computer-vision tool that digitizes Japanese MEP engineering drawings into structured Excel.

Built the OCR and alignment core: a resilient dual-engine pipeline (Azure Document Intelligence + PaddleOCR) hardened against noisy, rotated scans, plus a geometric graph-based algorithm that maps each text label to its corresponding pipe segment — including complex multi-size leader-line callouts.

PythonAzure Document IntelligencePaddleOCROpenCV

Details

code-distilling

Open source

A Claude Code / Codex plugin for disciplined feature porting from reference repos.

A five-stage workflow (analysis → spec → plan → implementation → gap-report) with human review gates, codifying how to extract the encoded decisions from a reference implementation and re-express them in your own project.

TypeScriptClaude CodeCodex

GitHub

Vision Assistant Device

Thesis

AI-powered assistive navigation device for visually impaired users.

Real-time object detection, obstacle avoidance, and voice-guided navigation using YOLO-World and speech recognition on an Intel RealSense depth camera.

PythonPyTorchYOLO-WorldIntel RealSense

GitHub

Digital Novel Platform

Side project

Novel reading and management platform with cloud asset storage.

Bookmarking, reading history, and cloud asset storage built with TypeScript, MongoDB, and Amazon S3.

TypeScriptMongoDBAmazon S3

GitHub

// skills

Languages

PythonGoRustTypeScriptC++SQL

LLM & Agents

LangGraphLangChainClaude Agent SDKGoogle ADKMCPAG-UICrewAIn8n

AI & ML

PyTorchTensorFlowOpenCVScikit-learnNumPy

Cloud & Infra

GCP (Vertex AI, Pub/Sub)AWS (SageMaker, Lambda, Bedrock, S3)DockergRPC

Data & Backend

PostgreSQL / pgvectorRedisChromaDBLanceDBFastAPITauriNext.js

// education

B.Eng, Computer Engineering

Ho Chi Minh City University of Technology (HCMUT) · 2020 – 2024

Thesis: Vision Assistant Device — assistive navigation for visually impaired users.

Certifications

•TOEIC 875/990 (2022)
•Coursera Specializations: Machine Learning, Deep Learning, NLP, Generative AI with LLMs (2023)

// contact

Let's build something.

Available for freelance & contract work — reach out and I'll get back to you quickly.

GitHub LinkedIn Email