Hung-Ta Chen

Hung-Ta Chen

Software Development Engineer at Amazon

I'm a software engineer in Amazon Stores, on the Seller Fees team within Selling Partner Services, where I build LLM-powered fee explanations and an AI fee agent for sellers. I have an M.S. in Statistics from UC Davis and a background in electrical and computer engineering, so I like work where backend engineering meets data and machine learning. Based in Seattle.

Get in touch
Portrait of Hung-Ta Chen

Experience

  1. Dec 2024 – Present

    Seattle, WA

    Amazon

    Software Development Engineer, Stores – Selling Partner Services (Seller Fees)

    LLM Fee Explainer

    Fee Explainer Introduction

    Turns a seller’s fee charge into a plain-language explanation across Manage Your Payments, Revenue Calculator and Seller Assistant. It serves ~625K explanations a day, with 38 fee types live across 29 marketplaces. I own it end to end:

    • Control plane: Designed a config-driven platform, with configuration in DynamoDB managed through an internal portal, so a new fee type onboards with configuration only instead of ~4 engineer-weeks of code. Replacing ~36 per-fee prompts with one cacheable template cut serving cost 55–58% per region (~$78K/yr) at ~81% cache hit.
    • Fee-change workflow: Built an AWS Step Functions workflow that detects fee changes, turns raw rate tables into explainable formulas with a symbolic engine, generates the explanation with an LLM, and runs an AI review agent before human approval and a multi-region rollout. 517 rate-card changes have gone live automatically (94% AI-approved), taking per-change engineering time from 0.5–3 hours to zero. The new O(K·N log N) algorithm replaced an O(N²) one, so a 1,260-row table that used to time out now finishes in ~9 minutes.
    • Models and reliability: Moved Nova Pro to Haiku 4.5 after an A/B test on 74 real prompts (100% vs 74% quality). Then moved to Claude Sonnet 5.5 with multi-model fallback and circuit breakers. Bedrock cost fell 80–83% (~$71–85K/yr), p90 latency fell 44–47% in every region, and no region has exceeded 0.1% daily faults since the fallback (it was up to 3.1% before).
    • India expansion: As the sole engineer, brought the Explainer to a marketplace whose fee structure differs from every other region. 10 fee types shipped, reaching 17,217 sellers in 30 days.

    Fee Agent

    About Seller Assistant

    An AI assistant that answers sellers’ fee questions. I own its model and experiment stack:

    • Moved it to Claude Sonnet 5.5 in six days behind go/no-go gates, then A/B-tested a low-effort setting. LLM-call p90/p99 fell from 18.3s/29.0s to 10.2s/13.2s, Bedrock cost fell 48% a day (74% per request) while traffic roughly doubled, and truncated answers dropped from 6.0% to 3.6%.
    • Built an experiment platform with dynamic model selection from Amazon Bedrock, so new models become available to test automatically, with promote/restore controls and review-gated production changes.

    Also

    • Built a daily pipeline-health report now used by 11 Fees teams, with all 64 of our team’s production pipelines meeting the ≤2-day time-to-prod goal (team average 0.1 day, down from 2.5 days in August).
    • Separately, led a zero-downtime migration of a production fleet to a new AWS region, moving 136M+ DynamoDB items and ~6TB of S3 data and eliminating $9.5K/month in AWS cost.
    JavaTypeScriptPythonAWSAmazon BedrockLLMs

  2. Oct 2024 – Nov 2024

    Remote (San Francisco, CA)

    Cielara AInow Causal Dynamics Lab

    Software Engineer

    • Built backend services in Go to scan AWS and GCP cloud resources and model configuration data into domain schemas for downstream security and analytics use cases.
    • Improved scan throughput by ~40% with goroutine-based parallelization and batch fetching, cutting end-to-end latency in large multi-account environments.
    GoAWSGCP

  3. May 2024 – Aug 2024

    Remote (New York, NY)

    Legion AI

    Software Engineer Intern

    • Built a Flask-based AI agent service using the OpenAI API and MetaGPT, integrated into a NestJS chatbot backend with Redis caching for low-latency, context-aware analytics conversations.
    • Built a RAG pipeline with LlamaIndex over a Pgvector database so the code-generation agent could fetch relevant examples on demand, improving reliability.
    FlaskNestJSRedisLlamaIndexDocker

Education

  • Sep 2022 – Jun 2024

    Davis, CA

    University of California, Davis

    M.S. Statistics, Data Science Track

  • Sep 2019 – Jan 2022

    Hsinchu, Taiwan

    National Chiao Tung University

    B.S. Electrical & Computer Engineering

  • Sep 2017 – Jun 2019

    Taipei, Taiwan

    National Taiwan University

    B.S. Biochemical Science & Technology

Projects

Skills

AI & LLM
  • Amazon Bedrock
  • LangGraph
  • LangChain
  • LlamaIndex
  • MCP
  • PyTorch
  • scikit-learn
Cloud
  • AWS
  • Lambda
  • ECS
  • Step Functions
  • S3
  • Aurora
  • CDK
  • Docker
Data
  • DynamoDB
  • PostgreSQL
  • Redis
  • Elasticsearch
  • Spark
  • SQL
Backend
  • Java
  • Spring Boot
  • Python
  • Flask
  • Django
  • Go
  • Scala
  • Node.js
Frontend
  • TypeScript
  • React
  • Next.js
  • Tailwind CSS

Let's talk

Open to interesting problems and good conversations. Send a message here or email me directly.

hungtachen0121@gmail.com