AI/ML Engineer • Seoul, South Korea

Hi, I'm Rohit Thakur

I build production-grade AI systems—from LLM-powered agents and real-time computer vision to scalable manufacturing analytics platforms serving global enterprises.

8+
Years Experience
15+
Production Systems
3
Publications
210+
ML Tests Built
GitHub Profile Get in Touch
About Me

Engineering AI for the Real World

I'm an AI/ML Engineer specializing in building intelligent systems that solve complex manufacturing and industrial challenges. My work spans the full AI stack—from designing LLM-powered conversational agents and predictive ML pipelines to deploying real-time computer vision systems on edge devices.

Currently at eMoldino, I architect a manufacturing intelligence ecosystem integrating AWS Bedrock, Snowflake, and FastAPI that serves global clients like PACCAR, Jaguar, and Eaton. Previously, I led an AI engineering team at OpenSysNet and developed ADAS/DMS systems at Pittasoft.

I hold an M.S. from GIST (South Korea) with a specialization in AI/ML, and I'm passionate about bridging research and production to deliver measurable business impact.

🤖

LLM Agent Architecture

Production agents with LangGraph, Bedrock, MCP

👁

Computer Vision

ADAS, DMS, industrial detection at 99.5% mAP

📊

Manufacturing Analytics

Capacity, downtime, scrap prediction pipelines

🎯

MLOps & Deployment

Docker, K8s, MLFlow, CI/CD, Langfuse

Technical Skills

Tools & Technologies

A comprehensive stack built across 8+ years of production AI development.

💻 Programming

PythonC++SQLRJavaJavaScript

🧠 AI/ML Frameworks

TensorFlowPyTorchKerasscikit-learnLangChainLangGraphDeepEval

👀 Computer Vision

OpenCVONNXMediaPipeYOLOv5/v8U-NetFFmpeg

✨ LLM & GenAI

AWS BedrockClaude 3.7GPT-4oRAGMCP ProtocolPrompt Engineering

☁ Data & Cloud

SnowflakeSnowparkPandasNumPySciPySparkKafkaAzure MLDatabricks

🛠 MLOps & DevOps

DockerKubernetesMLFlowGitCI/CDLangfuseFastAPIStreamlit
Experience

Career Journey

Building AI solutions across manufacturing, automotive, and IoT domains.

AI-ML Engineer
eMoldino
Seoul, South Korea
Apr 2025 – Present
Spearheaded the design and delivery of eMoldino's AI analytics platform for B2B injection molding manufacturing, building 15+ production-grade systems spanning LLM-powered agents, predictive ML pipelines, real-time IoT analytics, and automated reporting tools. Architected a full-stack manufacturing intelligence ecosystem integrating AWS Bedrock, Snowflake, and FastAPI serving multiple global clients (PACCAR, Jaguar, Eaton).
  • Built a Production Insight LLM Agent using AWS Bedrock (Claude 3.7 Sonnet), LangGraph, and Snowflake with intent-driven routing and a React frontend with real-time streaming
  • Developed AI-Powered Cycle Time Deviation Analysis with adaptive window algorithms (30-90 day) and statistical significance testing (Mann-Whitney U, Chi-square)
  • Deployed Scrap Rate Prediction pipeline improving test MAE by 33% (16.3% → 10.9%)
  • Engineered Manufacturing MCP Server (v3.1) with 4 core analytics tools via FastAPI with async job queues and multi-user session support
  • Created Combined Tooling Report Agent for executive summaries and actionable recommendations using LangChain and AWS Bedrock
  • Built LLM Evaluation Framework with 210+ tests, LLM-as-Judge and DeepEval metrics across 4 manufacturing modules
  • Designed Capacity Risk Analysis with Physics-of-Loss methodology, rolling risk scores, and OLS-based burn-up forecasting with Langfuse tracing
  • Developed Downtime Detection system extracting 11+ engineered features per event with automated Snowflake upload
  • Built acceleration sensor pipelines with RMS/Crest Factor metrics and KS testing for anomaly detection
AI Engineering Manager
OpenSysNet Co., Ltd.
Pangyo, Seoul, South Korea
May 2022 – Apr 2025
  • Led and mentored a team of 5 AI engineers, implementing SAFe Agile to improve delivery
  • Pantograph Monitoring System: real-time railway fault detection at 95.5% accuracy using YOLOv8 and edge processing
  • Sorting Robot Vision System: 99.5% mAP@50 object detection, improving manufacturing efficiency by 20%
  • LSTM and Gradient Boosting models for energy consumption optimization achieving 12.3% energy savings
  • Computer vision applications: golf swing analysis, face/fire/hand detection, drone OCR
Research & Development Engineer
Pittasoft Co., Ltd.
Seoul, South Korea
Sep 2017 – Apr 2022
  • Led ADAS development: forward vehicle detection, traffic light recognition, Forward Vehicle Start Alarm
  • Trained TensorFlow YOLO models for Driver Monitoring System (DMS) for distraction and drowsiness detection
  • Applied VGG-based U-Net for high-accuracy medical image segmentation
  • Built end-to-end video processing pipelines using OpenCV and FFmpeg for real-time streams
Key Projects

What I've Built

Highlights from my manufacturing AI and computer vision work.

Production Insight LLM Agent

v1.8 • Full-Stack

Collaborative LLM agent for manufacturing IoT data with intent-driven routing, hierarchical graph workflows, knowledge graph integration, and real-time streaming React frontend.

AWS BedrockLangGraphFastAPIReactSnowflakeMCP

Manufacturing MCP Server

v3.1 • 4 Analytics Tools

B2B analytics server providing ROI Analysis, RunRate, Capacity Planning, and CT Efficiency via MCP protocol with async job queue, SQLite persistence, and rate limiting.

FastAPISnowparkSQLAlchemyBedrockPydantic

LLM Evaluation Framework

210+ Tests • 4 Modules

End-to-end evaluation pipeline comparing prompt variants with LLM-as-Judge scoring (6 criteria) and DeepEval metrics (Faithfulness, Hallucination, Answer Relevancy).

DeepEvalClaudeGPT-4oSnowflakepytest

Capacity Risk Analysis

Physics-of-Loss • AI Insights

Interactive dashboard with rolling risk scores, waterfall analysis, OLS forecasting, and Bedrock-powered AI insights with Langfuse observability tracing.

StreamlitPlotlyBedrockLangfuseSnowflake

Pantograph Monitoring System

95.5% Accuracy • Award Winner

Real-time railway pantograph fault detection system using YOLOv8 with edge processing for live video streams. Won Innovation Award at OpenSysNet.

YOLOv8OpenCVEdge AIPython

Sorting Robot Vision

99.5% mAP@50 • +20% Efficiency

AI object detection system for industrial sorting robots that improved manufacturing efficiency by 20% with near-perfect detection accuracy.

YOLOOpenCVONNXPython
Research

Publications

Energy Consumption Prediction Using LSTM Models

International Conference on Machine Learning and Applications (ICMLA)

Concurrent Food Localization and Recognition

Korean Institute of Communications and Information Sciences (KICS), 2017

Augmentation of RGB Dataset for Improved Performance on Infrared Classifier

Electronics and Information Communications Conference, 2016

Education

Academic Background

M.S. in Electrical Engineering & Computer Science

Gwangju Institute of Science and Technology (GIST), South Korea • 2017
Specialization: Artificial Intelligence and Machine Learning

B.Tech. in Electronics & Communication Engineering

Green Hills Engineering College, India • 2014

Get in Touch

Let's Connect

Open to collaborations, consulting, and new opportunities in AI/ML.

✉

Email

trohit920@gmail.com
💻

GitHub

trohit920
📱

Phone

+82-10-2051-5650
🌏

Location

Seoul, South Korea