AI Research Engineer

Hanjun Cho

Research

(RAG, Retrieval, Evaluation, Numrical Reasoning, Table QA)

  • NeurIPS Reviewer 2026
  • Hanjun Cho, Jay-Yoon Lee. "RARE: Redundancy-Aware Retrieval Evaluation Framework for High-Similarity Corpora." ACL 2026.
  • Hanjun Cho, Gahyun Yoo, Hanseong Kim, Jay-Yoon Lee. "Generalizing Numerical Reasoning in Table Data through Operation Sketches and Self-Supervised Learning." TACL 2026.

Open-Source Contribution

(Quantization, LLM Inference Optimization)

  • NVIDIA TensorRT-LLM: Support Qwen3 Dense & MoE TensorRT compile engines.
  • vLLM: fixing FP8/NVFP4 quantization bug for sequence classification models.

Engineering

(System, Backend, Infra)

  • Multi-tenant agentic AI platform for financial institution, deployed as a hybrid on-premise architecture (network-segregated). (installation & ops)
  • Single-tenant RAG on AWS EKS (K8s) for enterprise client: 300 RPS in production, validated up to 10K RPS in loads. (deploy & ops)
  • GraphRAG over interconnected enterprise documents for portfolio-level strategic queries. (deploy)
  • Multi-Region embedding inference servers across KR/US/JP, 30M req/month. (deploy & ops)
  • TensorRT Engine & Triton Server-based priority/distributed scheduling system for inference serving: (deploy & ops)
    • Reranker: Top-3 accuracy +4–21%; latency 3.3s → 0.3s
    • Embedding: Top-3 accuracy +~20%; latency 2000ms → 120ms
  • Prometheus & Grafana based monitoring system (deploy & ops)

Others

(Leadership, Product)

  • Core Part Lead: Led the Core team (3 engineers, ~6 months) driving technical differentiation: OCR & VL embedding serving (63% lower latency), Agentic RAG (accuracy 45% → 82%).
  • Ralli: enterprise RAG product, sole developer: built end-to-end in ~2 months, shipped to first paying customer (Hyundai Motor Securities).