Skip to content

Selected work

Systems shipped to production, not slideware.

A close look at four projects from the last 18 months — the architecture decisions, the measured results, and the scope of ownership behind each.

01Multi-Agent SaaS

AI Learning-Management Platform

Role
Lead backend / platform engineer
Period
Late 2025 — Present

1,300+ authored commits on the core backend

A multi-tenant SaaS platform automating structured learning programs and their member community — course management, progress tracking, AI-assisted grading, and conversational student support.

  • Built a from-scratch multi-agent AI engine (router → planner → executor) on a hand-rolled multi-model LLM client — streaming, multimodal input, prompt caching, model fallback, and cost-optimized routing via OpenRouter (Anthropic, Gemini, OpenAI).
  • Shipped AI grading with a human-in-the-loop review gate: rubric-based scoring (0–100), score-gated final assignments, and professor review states auto-posted back to the community layer.
  • Engineered a custom RAG pipeline (multi-strategy chunking over pgvector, hybrid search) and data synchronization keeping three integrated systems consistent via 20+ sync tools.
  • Delivered a hardened white-label variant with JWKS auth and organization-level data isolation as an independently deployable mirror.
FastAPIOpenRouterpgvectorSupabaseRedisSSEStripeVercel
02Full-Stack + Agentic

Multi-Tenant Travelling Assistant

Role
Full-stack engineer (front-end lead)
Period
Late 2024 — Present

1,500+ commits · single-pass hybrid retrieval hits 0.97 facet recall, 3.8–4.0× faster than agentic alternatives

A two-sided consumer + business platform where organizations publish rich experiences and users discover them, chat with an AI companion, and generate personal stories — with a full CMS and analytics dashboard.

  • Built the Next.js App Router front-end with organization-scoped routing, role-based auth, and streaming AI UX — live reasoning trails, TTS playback, and voice-to-text capture.
  • Delivered a LlamaIndex agentic backend with a hybrid routing layer: rule-based multilingual fast-path plus an LLM router deciding web search vs. knowledge lookup vs. weather.
  • Built hybrid retrieval — FAISS + BM25, ~1K-token chunks, Gemini + VoyageAI embeddings — inside a RAG enrichment workspace with live research-progress streaming.
  • Shipped 7-locale i18n (EN, FR, JA, KO, RU, VI, ZH) on a Redux Toolkit / RTK Query data layer with a broad Playwright + Vitest suite focused on tenant isolation.
Next.jsFastAPILlamaIndexFAISSSupabaseRedux Toolkit
03Regulated Health-Tech

Occupational Health & Longevity Analytics

Role
Senior full-stack / AI engineer
Period
Q2 2026 — Present

200+ authored commits · aggregate-only privacy model (N ≥ 30)

A family of regulated health-tech products: a personalized longevity coaching app with multi-domain risk scoring, and a workplace musculoskeletal-risk intelligence platform.

  • Built a multi-domain clinical risk engine (TypeScript) scoring six domains — cardiovascular, metabolic, musculoskeletal, neurodegenerative, oncological, and biological age — with per-domain regression tests and snapshot suites for numerical stability.
  • Designed the AI assistant layer with DB-stored persona agents (clinician, nutrition, PT, supplement-advisor) exposed via Vercel AI SDK + MCP tool-calling.
  • Shipped the Atlas MSK chatbot end-to-end: intent classifier (Gemini Flash) + Claude synthesis, cache-first + MSSQL fan-out query layer, and pgvector RAG running in parallel.
  • Owned security & compliance hardening: PII redaction at the data-view layer, Postgres RLS, an LLM-as-judge evaluation harness with per-persona suites, and medical-device risk documentation.
Next.js 16TypeScriptVercel AI SDKMCPSupabaseMSSQL
04Developer Tooling / MCP

Claude Code Agent Orchestration Plugin

Role
Creator / primary author
Period
Q1 2026 — Present

130+ authored commits across template + plugins

A distributable plugin and agent-template system for Claude Code / Claude Desktop that standardizes AI-assisted software delivery for a whole team — an 8-role agent org, a skill library, and session lifecycle hooks.

  • Designed an 8-role agent template library (PM, Developer, QA, DevOps, Designer, Writer, Publisher, Marketer) backed by rich, standalone skill files.
  • Authored a universal orchestrator (assess → plan → distribute → collect → validate → report) that spawns waves of parallel sub-agents with handoff-document contracts.
  • Packaged it as a real plugin with auto-install & sync on session start, version tracking via the GitHub Releases API, and consent-first update notifications.
  • Built a delivery-telemetry subsystem capturing LLM token usage and human-hour records, plus MCP gateways exposing curated CRUD/agent tools. Adopted by clients across the US, Australia, and Vietnam.
PythonTypeScriptMCPFastMCPClaude Code
—

Experience

Edge8 AI

AI Automation Engineer

Jan 2025 — Present · Remote

  • Architected and built multiple multi-tenant SaaS products and internal platforms across AI-native, content, analytics, and regulated health-tech domains — all on Supabase (PostgreSQL) and Vercel.
  • Designed production multi-agent AI systems with custom multi-model LLM clients, model fallback, cost-optimized routing, and streaming SSE.
  • Delivered real-time AI UX (streaming reasoning, TTS/STT), 7-locale i18n, Stripe billing, and RAG pipelines with hybrid search (pgvector, FAISS, BM25).
  • Enforced engineering quality: unit/integration/E2E tests (Vitest, Pytest, Playwright), CI/CD, tenant-isolation suites, and LLM evaluation harnesses.

University of Ulsan, South Korea

Research Assistant

Sep 2022 — Aug 2024 · Ulsan, KR

  • Led deep learning research on error correction coding (ECC); published first-author in IEEE Trans. Cogn. Commun. Netw. and won Best Paper Award at ICGHIT 2024.
  • Designed Transformer, CNN, and RNN architectures in PyTorch, TensorFlow, and Keras for wireless communication systems.

Bosch Global Software Vietnam

Embedded Software Engineer

Jun 2021 — Aug 2022 · Ho Chi Minh City, VN

  • Developed and validated automotive embedded software for Engine Control Units (ECU) across the full SDLC; programmed in C, Python, and XML with CAN/FlexRay protocols.
  • Delivered features under Agile Scrum and V-Model; maintained CI/CD pipelines on a Jenkins-based platform.