Automation on CZ Builds

I Built a Fully Automated AI Newsletter in 3 Days. Here's Why I'm Pivoting.

Sun, 12 Apr 2026 00:00:00 +0000

Last week I built PandaBrief — a fully automated daily newsletter that translates Chinese tech news into English. From RSS ingestion to LLM scoring to email delivery, zero human intervention required.

The entire pipeline runs on GitHub Actions, costs $0/month, and handles everything from content curation to subscriber management to self-healing code fixes.

Then I realized the product itself might not be worth building.

What I Built

The system has 8 automated steps that run every day:

RSS Fetch (16 sources) → LLM Scoring → Dedup → Translation → 
Quality Check → Render HTML → Send Email → Operator Report

Tech stack:

Content pipeline: Python + Groq/Gemini API (free tier, with auto-fallback)
Hosting: Cloudflare Pages + KV storage (free)
Email: Resend (free tier)
CI/CD: GitHub Actions (free)
Auto-fix: LLM reads GitHub Issues → generates patches → opens PRs → emails me to review

Total monthly cost: $0

The Interesting Engineering Problems

LLM rate limits: Groq’s free tier gives you 100K tokens/day. My pipeline burned through that in one run. Solution: built a provider fallback chain (Groq → Gemini → Zhipu) that automatically switches when one provider hits its limit.

Chinese character corruption: When asking the LLM to output complete file contents for auto-fix PRs, it would corrupt Chinese characters in source configs (量子位 → 量种体). Solution: switched from “output the whole file” to “output search-and-replace patches” — the LLM copies exact text and only modifies the target section.

Quality gate false positives: The translation step uses full article text (3000 chars), but the quality checker only saw a 200-char snippet. It flagged 9/10 translations as “hallucinated” because it couldn’t find details that were actually in the full article. Solution: give the checker more context and tell it “the translator had access to the full article.”

Why I’m Pivoting

The engineering was fun. But the product has fundamental problems:

Translation is no longer a moat. In 2026, anyone can paste Chinese text into ChatGPT. A newsletter that translates is competing with a free, instant, universal tool.
The audience doesn’t have a strong pain point. “Staying updated on Chinese tech” is nice-to-have for Western devs, not must-have. Nobody’s workflow breaks without it.
AI Overviews are eating informational content. Google’s AI Overviews now reduce click-through rates by 58% for informational queries. The trend is accelerating.
The Chinese AI narrative has a credibility problem. Model distillation controversies, benchmark optimization concerns — the Western dev community is increasingly skeptical. Translating these stories doesn’t add trust, it just amplifies noise.

What I Actually Learned

The project wasn’t a waste. I learned more in 3 days of building than in weeks of tutorials:

How to orchestrate multi-step LLM pipelines with fallback and retry logic
Cloudflare Pages + KV as a zero-cost deployment platform
GitHub Actions as a free cron + CI/CD system
How to make LLMs generate safe, reviewable code changes (search-and-replace, not full rewrites)

Most importantly: the ability to quickly build and ship automated systems is valuable, but it needs to be pointed at a real problem.

What’s Next

I’m an IC (integrated circuit) student with FPGA experience. The AI arms race everyone talks about is fundamentally a hardware arms race. I’m going to explore the intersection of AI and hardware — starting with running neural network inference on my Zynq-7020 FPGA.

I’ll document everything here. Follow along if you’re interested in where silicon meets intelligence.

This is the first post on this blog. I’m building in public — sharing what I learn, build, and think about at the intersection of AI and hardware.

I Built a Personal AI News Agent That Emails Me Every Morning

Sun, 12 Apr 2026 00:00:00 +0000

I spend too much time scrolling Hacker News, arxiv, and tech Twitter. Most of it is noise. So I built a system that does the filtering for me — every morning at 6 AM, I get an email with the 10 most relevant stories across AI, semiconductors, and space.

Fully automated. Zero cost. Built in an afternoon.

The Problem

My information diet was broken:

Douyin/YouTube recommendations — optimized for engagement, not relevance
Manually checking 10+ sites — time-consuming, easy to miss things
Asking ChatGPT — no context about what I specifically care about

I wanted something that knows my interests (AI hardware, FPGA, space engineering) and filters 200+ daily articles down to the 10 that actually matter.

How It Works

9 RSS Sources → Fetch → LLM Scoring → Summarize (bilingual) → Email

Every day at 6 AM Beijing time, a GitHub Actions workflow:

Fetches ~150 articles from 9 sources (Hacker News, arxiv, Ars Technica Space, SemiAnalysis, EE Times, etc.)
Scores each article 1-10 based on my specific interests using Gemini
Summarizes the top 10 in both English and Chinese
Emails the digest to me with links to every original article

The Scoring Prompt

This is where the personalization lives. The LLM knows I’m an IC student interested in:

AI inference on edge hardware (FPGA/ASIC)
Semiconductor industry moves
Space technology engineering details
Open source AI tools
Contrarian technical takes

A paper about efficient transformer inference on resource-constrained chips scores 9/10. A generic “AI will change everything” opinion piece scores 2/10.

Model Fallback Chain

Free API tiers have rate limits. Instead of failing when one model is exhausted, the system automatically falls back:

gemini-2.5-flash (best quality, 20 req/day)
 ↓ exhausted? auto-switch
gemini-2.0-flash (good quality, 1500 req/day)
 ↓ exhausted? auto-switch
gemini-2.0-flash-lite (basic, unlimited)

The most important items get processed by the best model. Later items gracefully degrade.

Sources

Source	Category	Why
Hacker News	AI & Tech	High signal-to-noise, community-validated
arxiv cs.AI	AI Research	Primary source for papers
arxiv cs.AR	Hardware	Computer architecture papers
SemiAnalysis	Hardware	Best semiconductor industry analysis
EE Times	Hardware	Electronics engineering news
MIT News AI	AI Research	Research breakthroughs
Ars Technica Space	Space	Eric Berger’s space reporting
Everyday Astronaut	Space	Deep engineering explanations
SpaceNews	Space	Industry coverage

Tech Stack

Pipeline: Python (feedparser + google-genai + resend)
Database: SQLite (ephemeral, rebuilt each run)
Scheduling: GitHub Actions cron (free)
LLM: Gemini API (free tier)
Email: Resend (free tier)
Total cost: $0/month

What I Learned from PandaBrief

This project reuses the architecture from PandaBrief, my previous newsletter project. Key improvements:

Model fallback chain — never fail because of rate limits
Bilingual output — Chinese + English in one prompt call
Personalized scoring — not generic “is this important” but “is this important to me”
Private by default — this is a personal tool, not a product

The biggest lesson: the same automation pipeline that couldn’t sustain a public product works perfectly as a personal productivity tool. The difference is that I’m the only user, so I can tune the scoring to exactly what I care about.

Try It Yourself

The code is at github.com/cz-builds/daily-digest. Fork it, change sources.yaml and the scoring prompt to match your interests, add your API keys as GitHub Secrets, and you’ll have your own AI-curated morning briefing.

This is part of my build in public series. I’m an IC student exploring the intersection of AI and hardware.