AI Engineer

Bibek Chaudhary

I'm an AI Engineer who turns LLMs into dependable, production-grade systems.

I've shipped agentic systems end to end — a slide-generation agent for a consumer AI assistant and a government-proposal agent that helped win a real contract. Beyond LLMs, I've worked on OCR, recommendation, and on-device model compression — and I compete as a Kaggle Competitions Master.

Experience

Tap any role to see the full story.

  1. 2023 — Now

    ESTsoft Current

    AI Research Engineer · Oct 2023 — Present · Seoul

    Designed and shipped an agentic system for end-to-end slide generation, turning researched content into visually consistent infographic presentations for Alan, ESTsoft's consumer AI assistant.
    • Hardened reliability with structured outputs, validation, parallel execution, retries, model fallbacks, and LangSmith tracing — so individual failures never interrupt a full generation job.
    • Built an evidence-grounded Notion retrieval system (query reformulation, multi-hop retrieval, coverage checks, citations, refusal rules), improving accuracy from 80% to ~92% on 30 complex queries.
    • Shipped a company-wide agent plugin for government proposal writing — eligibility screening, evidence collection, technical planning, budget validation, drafting, and compliance review with human approval gates. Used on real submissions: five passed screening, one won a contract.
    • Developed a TypeScript/Node.js library that converts Markdown into Korean HWPX templates while preserving layout compatibility, eliminating per-template configuration.
    • Agentic Systems
    • RAG
    • LLM Evaluation
    • LangGraph
    • TypeScript
  2. 2022 — 2023

    Upstage

    AI Research Engineer · Jan 2022 — Sep 2023 · Seoul

    Trained diverse text-detection models for an ensemble that won the 2023 ICDAR Hierarchical Text Detection competition.
    • Built training and inference pipelines for OCR models on Korean financial documents, enabling automated extraction from complex document layouts.
    • Improved recommendation quality with collaborative-filtering features and enhanced negative down-sampling, strengthening ranking performance.
    • Automated internal knowledge capture with a GPT-based workflow that summarized Slack discussions into organized Notion pages.
    • OCR
    • Text Detection
    • Recommendation
    • Ensembles
  3. 2019 — 2021

    Nota

    AI Research Engineer · Jan 2019 — Dec 2021 · Seoul

    Owned the pipeline integrating pruning, knowledge distillation, and quantization into Netspresso, Nota's production model-compression platform.
    • Implemented and trained benchmark models across classification, segmentation, and detection — ResNet, MobileNet, U-Net, YOLO, SSD — to validate Netspresso across CV tasks.
    • Benchmarked accuracy and edge performance before and after compression with TensorRT and ONNX, validating deployment on NVIDIA Jetson Nano and Raspberry Pi 4.
    • Led a three-person research team tracking object-detection advances and translating research into working implementations for benchmarking.
    • Model Compression
    • Quantization
    • Edge AI
    • Computer Vision

Side Projects

Competitive machine learning on Kaggle.

  1. 2019 — 2023

    Kaggle Competitions Master

    Global rank 5,251 / 210,363 · 1 gold, 4 silver, 4 bronze

    Finished 12th of 1,588 teams in Happywhale — Whale & Dolphin Identification, earning a gold medal for large-scale, metric-learning-based individual re-identification.
    • 16th / 2,574 — OTTO Multi-Objective Recommender System (session-based e-commerce recommendation).
    • 17th / 2,025 — TensorFlow: Help Protect the Great Barrier Reef (underwater object detection).
    • 26th / 1,531 — Understanding Clouds from Satellite Images (semantic segmentation).
    • 49th / 801 — BirdCLEF 2022 (audio bird-call identification).
    • Nine competition medals across 24 competitions — top ~3% of over 210,000 competitors worldwide.
    • Metric Learning
    • Computer Vision
    • Segmentation
    • Recommendation
    • Audio
    • PyTorch

Education

  1. 2015 — 2019