Jieyi
Mount Rainier across Puget Sound, seen from the Seattle waterfront.

Jieyi Deng

Data Scientist,Growth, Product & Agentic Analytics

I’m a Data Scientist who takes ambiguous growth and product problems from framing through analytics or machine learning and into product and engineering delivery. My work spans risk prediction, recommendation, growth measurement and agentic analytics.

I like problems that arrive unstructured, where the first job is deciding what the question actually is, and the second is finding out whether the data can answer it.

Projects

  • Pallas Ads

    github.com/Jieyi-Deng/Pallas-ads

    A free, open-source advertising analytics plugin for Codex and Claude Code. It analyzes Google, Meta and TikTok Ads data through supported account connections or CSV/XLSX imports, preserving business context and producing traceable reports and local dashboards.

    • Advertising analytics

  • Agentic Knowledge Vault

    github.com/Jieyi-Deng/agentic-knowledge-vault

    A provenance-first knowledge system for people and agents. It keeps source captures immutable, links every processed note back to its evidence and uses structured metadata to make retrieval reliable.

    • Agent workflow

  • Jieyi Skills

    github.com/Jieyi-Deng/jieyi-skills

    Executable agent workflows for capturing sources and turning vetted notes into offline study material. The scripts standardize the steps that should remain deterministic rather than improvised.

    • SKILL.md

Writing

All writing

  • Pallas Ads: Read the Change Before You Act on It

    An account's cost per click can rise while every campaign gets cheaper. How Pallas separates what changed from why, keeps the counting in tested code and ties ad metrics back to business value.

    6 min read

    Hypothetical two-campaign example: both campaign CPCs fall, but the higher-CPC Campaign B's click share rises from 10% to 50%, lifting account CPC from $0.650 to $1.175. Campaign A is grey; Campaign B is blue.

  • More Than Prompting

    SKILL.md turns a prompt into an operational protocol. What that looks like in advertising analytics, where a fluent answer built on the wrong metric is worse than no answer at all.

    4 min read

    Sequence diagram of the analysis skill, with lanes for the user, the agent, the insights skill runtime and the ads data backend, divided into phases for forming the query, validating and executing it, and interpreting the result.

  • My 2025 with ChatGPT: What a "Year in Review" Doesn't Show

    I exported a full year of ChatGPT history and analysed it. The conversation tree holds more than the official recap shows — including three different URL fields that reveal how retrieved sources actually reach an answer.

    3 min read

    Tree diagram of the exported conversation JSON, branching from the root object into title, conversation_id, mapping, timestamps and current_node, then down through node, message, content and metadata fields.

About

Jieyi Deng, in a black turtleneck against a plain grey background.

Born and raised in China and based in Seattle. Outside work, I’m usually taking photos, hiking or finding another reason to look toward Mount Rainier.

More about me