Enterprise Data & AI Architect | Hands-on AI Engineer

I design and build data and AI solutions that connect complex enterprise systems with real business needs.

Over the past 8 years, I've worked across enterprise data architecture, applied machine learning, AI systems, and customer-facing technical consulting — from understanding business requirements and evaluating technical options to building solutions and validating their impact.

At CJ CheilJedang, I've worked on enterprise data platform modernization, including hands-on validation of SAP Business Data Cloud–Databricks integration and Delta Sharing. My work involves evaluating architecture trade-offs across performance, cost, governance, and AI readiness. I've also developed AI solutions for procurement decision support, material classification, entity resolution, RAG, and knowledge graph-based profitability analysis.

Previously at Hansol PNS, I worked with clients across manufacturing, logistics, pharmaceuticals, and distribution, collaborating with sales teams on technical PoCs, solution architecture, and end-to-end project delivery.

I enjoy solving problems where the right technical approach isn't obvious at first — understanding the business context, designing practical solutions, building them hands-on, and testing whether they actually work.

Currently exploring AI agents, retrieval and evidence-grounded reasoning, evaluation, and reliable AI systems.

An AI agent that investigates production incidents across a synthetic microservice environment.

Uses native tool calling to inspect logs, metrics, distributed traces, deployments, configuration changes, dependencies, and runbooks.

Built fault-injection scenarios for deployment regressions, dependency failures, bad configurations, connection-pool exhaustion, and traffic spikes — with evidence validation, bounded investigation loops, human approval for high-risk remediation, and stateful recovery verification.

Python FastAPI LLM Tool Calling Distributed Tracing Agent Evaluation

A research-inspired fault-localization system based on the AutoFL approach.

Given a stack trace, ranks the functions most likely to contain the bug using AST-level code chunking, hybrid keyword + embedding retrieval, and optional Claude reranking.

Achieved 100% Top-1 / Top-5 retrieval accuracy on a 5-case controlled benchmark, with automated evaluation through GitHub Actions.

Python AST Embeddings Hybrid Retrieval Claude API Evaluation

A multilingual classification system for 100K+ records, combining learned embeddings with rule-based logic and human feedback.

Built with Triplet Loss + Hard Negative Mining and a human-in-the-loop retraining workflow, achieving 95%+ classification accuracy.

PyTorch Embeddings Triplet Loss Hard Negative Mining Human-in-the-loop

A real-time voice AI system for detecting and responding to suspicious calls.

Whisper handles speech-to-text, a lightweight rule-based layer provides low-latency and explainable risk detection, and GPT is used selectively for language generation.

Built end-to-end from backend and telephony integration to frontend and deployment.

Whisper OpenAI API Flask Twilio React GitHub Actions Render

A multilingual entity-resolution system for ~50K global records with inconsistent identifiers and noisy text.

Reduced a naïve ~151M pairwise comparison space to tens of thousands of candidates through:

Blocking → ANN Retrieval → Composite Similarity Scoring

Python ANN Embeddings Entity Resolution Multilingual NLP

An ML decision system that treats commodity purchasing as a sequential decision problem rather than just a forecasting problem.

Experiments with XGBoost, Temporal Fusion Transformer, and DQN to generate Buy / Split / Wait recommendations, exposed through a REST serving layer.

XGBoost TFT DQN Reinforcement Learning Time Series REST API

AI / ML
LLM APIs Tool Calling RAG Embeddings PyTorch scikit-learn Reinforcement Learning Whisper LangChain

Backend / Systems
Python FastAPI Flask Django Spring REST APIs SQL

Frontend
React TypeScript Vue.js

Cloud / Infrastructure
AWS Lambda API Gateway S3 GitHub Actions Render

Enterprise Systems
SAP Ariba BW/SAC CAP HANA

I'm interested in the space between a model that works in a notebook and an AI system that actually works.

That means thinking about more than model accuracy:

retrieval · evaluation · latency · grounding · failure modes · human feedback · system integration · deployment

I like experimenting with new AI techniques, but I'm even more interested in understanding when they work, when they fail, and how to turn them into reliable systems.