Eric Cao

After enough code, you start asking why systems are designed this way

Worked as a big data engineer, moved through Python full-stack, built LLM apps. But what truly fascinates me was never the technology itself — it's why a system is designed the way it is, where it quietly fails, and what work you can actually hand off to AI without regretting it.

01. About

My first job was in big data engineering — and what that shaped wasn't a specific skill, but a way of seeing systems: at sufficient scale, small problems become fundamental problems. What looks like a performance issue is often data quality; what looks like a logic bug is rooted in the data model. That instinct of 'dig one level deeper' has left its mark on everything I've done since.

Later I moved into LLM applications — and that's where my engineering intuition got completely overturned for the first time. Traditional software bugs throw errors. LLM 'bugs' are hallucinations: they look right, then you realize they're wrong, and you can't capture them with tests. Performance isn't latency, it's recall — things with no ground truth. Understanding an AI system requires a completely different set of engineering instincts.

Now I spend a lot of time on one question: where does enterprise AI adoption actually fail? The answer is counterintuitive — it's rarely the model. It's the absence of standardized processes. Layer AI onto a workflow with no clear inputs and outputs, and you only amplify the existing chaos. MIT and McKinsey research keeps arriving at the same conclusion; I've verified it in real projects too: AI generates measurable returns almost exclusively where the process was clarified first, and AI was introduced gradually — not the other way around.

02. Experience

Focused on Practical AI

2026 — Present

Enterprise AI adoption · workflow change

Reading source code, opening PRs, and asking what trade-offs drive the design decisions. The real challenges of enterprise AI live in the implementation details, not the slides.

AI Agent Claude Code Workflow

Python & Full-Stack Developer

2023 — 2025

Full-stack development · LLM apps

Moving from traditional software into LLM apps was the first moment I had to rebuild my engineering intuition from scratch.

Python TypeScript LLM React

Big Data Engineer

2021 — 2023

Jike (即刻)

At Jike, scale forced me to look deeper — what looked like a performance problem turned out to be data quality; what looked like a logic issue was rooted in the data model design.

Big Data Java Data Pipeline
03. Projects

Airflow extended API plugin—wrapping RESTful APIs via CLI commands for Airflow v2.X. More flexible DAG management and easier operations.

Python Airflow REST API CLI

One-click speed test and switch for pip & uv mirrors. Intelligently detects environment and auto-selects fastest mirror.

Python CLI Networking

A minimalist Mermaid diagram editor—no login required, real-time preview, export to PNG/SVG with sketch style support.

Diagram guides →
TypeScript React Mermaid Canvas
04. Thoughts

AI Agent in Production

No empty concepts or API tutorials—focus on real workflow pain points of AI Agents. From single-scenario validation to stable reuse, exploring long-running task stability, multi-tool flow design, and deep business adaptation.

Mar 2025

Big Data Engineering

No textbook concepts—focus on real business scenarios for big data pipeline setup, optimization and deployment. Full process practice from data collection, cleaning, storage to analytics.

Mar 2025

System & Performance Tuning

No theoretical parameters—focus on end-to-end performance tuning. From code optimization, architecture design to data pipeline acceleration and service stress testing.

Mar 2025

Philosophy of Tools

Good tools should be like air—invisible when present, immediately missed when absent. Not chasing tech trends, just polishing things that solve real problems. A tool's value isn't in its complexity, but in how much it frees its user.

Mar 2025

Workflow Automation

No scattered tool plugins—focus on how to use AI and automation to eliminate repetitive work. Breaking down the full efficiency workflow from requirements to implementation.

Mar 2025
05. Blog
Can an AI Get Smarter on Its Own, the Way an Intern Does?

Hermes Agent's pitch is 'the agent that grows with you' — an AI that gets smarter on its own, like an intern who learns. I took it literally: with the model frozen, can the agent make itself smarter just by writing its own skills? The answer is counter-intuitive — a human-written skill takes the same model from 10% to 74%, but the skill the agent writes itself closes less than half that gap.

2026-06-12
Every Commit Writes Just 2 Files: How lakeFS Brings Git to the Data Lake

An ETL job corrupts tens of thousands of parquet files, and when you open the S3 bucket you realize — there are no commits, no rollbacks, and the overwritten objects are gone forever. This article tears apart the design of lakeFS Graveler: how a two-level Merkle tree combined with hash-based chunking allows every commit at billion-file scale to write just 2 new files and reuse 99% of content in place — and how it shares the same mathematics as blockchain while pursuing the exact opposite goal.

2026-06-08
OpenClaw Isn't the Answer Yet: Reflections After Burning 300 Million Tokens in a Week

After a week of deep use and 300 million tokens burned, the verdict on OpenClaw: the problem isn't that AI isn't smart enough — it's that pure natural language interaction has fundamental flaws for execution-based tasks. An analysis of four interaction modes through the lens of one core question: who absorbs the ambiguity?

2026-03-10