AI/ML Engineer · Seoul, South Korea
Agentic LLM systems · Uncertainty & evaluation · Real-time edge vision
I build AI systems that survive production, and I spend most of my time on the part people skip: knowing when the model should be trusted.
That means agentic LLM applications with measured evaluation rather than vibes, real-time computer vision tuned to run on edge silicon, and the manufacturing analytics that connect a sensor reading to a business decision.
Currently at eMoldino, building a manufacturing intelligence platform on AWS Bedrock, LangGraph and Snowflake.
| Project | What it does |
|---|---|
| llm-uncertainty-quantification | Token- and sentence-level uncertainty for local LLMs. Semantic entropy via NLI clustering, calibration, risk-coverage, bootstrap CIs. Found a free single forward pass beating 10-sample semantic entropy, and traced exactly why. |
| CabinSense | Real-time in-cabin driver and occupant monitoring. ONNX face detection, 98-point landmarks, head-pose geometry and a multi-task classifier, INT8-quantized for SoC deployment. |
| Agentic RAG with Guardrails | Fully local agentic RAG with hybrid retrieval, NLI groundedness checks, prompt-injection defense and a judge-free evaluation harness. |
33% lower test MAE on scrap-rate prediction (16.3% -> 10.9%)
95.5% fault-detection accuracy, railway pantograph monitoring
99.5% mAP@50, sorting-robot vision pipeline
20% manufacturing efficiency gain
12.3% energy saved via anomaly detection and control
210+ automated LLM evaluation tests in production
LLM & GenAI
LangGraph LangChain Agentic RAG MCP AWS Bedrock DeepEval LLM-as-Judge Guardrails Hybrid retrieval FAISS Chroma
Vision & Edge
PyTorch TensorFlow OpenCV YOLOv5/v8 U-Net ONNX Runtime INT8 TensorRT MediaPipe
Data & Platform
Snowflake Snowpark Spark Kafka FastAPI Docker Kubernetes MLflow Langfuse Databricks
M.S. Electrical Engineering & Computer Science — Gwangju Institute of Science and Technology (GIST), 2017 · AI and machine learning
Published at ICMLA and KICS. Innovation Award at OpenSysNet, Employee of the Year at Pittasoft. IEEE member.
English (fluent) · Korean (professional working) · Hindi (native)
Open to AI/ML engineering roles — agentic systems, evaluation and reliability, or real-time vision.
trohit920.github.io