To all the engineers who got paged during the AWS outage last night: I feel for you. Being on-call is rough. Being on-call for something you can’t fix is existentially rough. You’re out here rebooting servers at 3 a.m., praying the cloud gods show mercy. Take a long nap today.

When I was learning Python, I kept forgetting everything. I’d ace a tutorial… then freeze the next day staring at a blank screen. What helped wasn’t more studying. It was more doing. Here's what I would tell myself if I could go back in time 👇

In the past 2 weeks, 85 people tried my Cantonese AI partner: 213 conversations 993 interactions A couple users came back 6–7 days straight Most dropped after day 1 Now I’m stuck: Do I keep fixing bugs + adding features? Or pause and set up Stripe to see if anyone pays?

I got a “fan message” for my AI prototype!! It wasn’t about the tech. It wasn’t about features. It said: “Thank YOU so much for making this. I love it. I will keep going.” 🥹

@ds-claudia.bsky.social explains why domain expertise often outweighs algorithmic complexity. In this Author Spotlight, she reflects on her path from corporate ML to freelance AI, mentoring newcomers, and the lessons learned along the way.

Woke up to see AI Weekender ranked #51 in Tech on Substack 🎉 I started on Jan 23 with one goal: publish every Thursday. Some posts hit, others sank. But showing up each week built momentum. Biggest lesson? Consistency compounds. aiweekender.substack.com

✨ In our latest Author Spotlight, @ds-claudia.bsky.social, who successfully shifted from a corporate role to freelance, provides hard-won advice on navigating career paths, mentoring newcomers, and building solutions you truly own.

Every project should either: - Build skills - Build assets - Build your brand If it’s doing none, stop.

As a solo data scientist, I’ve found 3 types of leverage: - Coaching gave me human leverage - Consulting gave me applied leverage - Content and AI tools? That’s scalable leverage Read more here: open.substack.com/pub/aiweeken...

I didn’t build my first RAG tool because I thought it would go viral. I built it because people kept DM’ing me asking how to break into data science or build with AI, all questions I’d already answered in my posts. Try it: assistant.ds-claudia.com

I was tired of AI tutorials that felt like science fair projects. So I built a real RAG project: it turns your Substack posts into a searchable AI assistant. This post walks through every step: from scraping to chunking to shipping. shorturl.at/GicNL

Building an AI product in 5 steps: 1. Have an idea 2. Prompt like a maniac 3. Watch it almost work 4. Panic, then debug 5. Ship it anyway Does this sound familiar, or is your chaos different?

Most of AI building isn’t about being smart. It’s about being stubborn enough to fix what breaks… And curious enough to try again.

Some projects take a weekend. Some take months. Both are valid. Both move you forward.

I don’t want to build flashy AI. I want to build stuff people actually use. aiweekender.substack.com/p/the-strate...

The real flex isn’t how smart your AI project is. It’s how clearly you can explain it to a non-technical friend.

Don’t aim for perfect. Aim for finished. Build momentum!

I don’t build AI projects to impress anyone. I build them to prove to myself I can. Why do you do what you do?

Most people are waiting for the confidence to begin. But confidence is what you earn after you start.

Your weekend build might not be world-changing. But it might change your world: Your confidence, Your direction, Your job.

I write AI technical tutorials not because I know everything, but because writing is how I figure things out.

Genuine question: Do you consider vibe coding with AI "cheating"? I see it like using: - Calculators instead of long division, or - GPS instead of memorizing maps, or - Driving a car instead of walking. Same destination, better tools. What's your perspective?

Cantonese diaspora: tired of sounding like a 5-year-old when talking to your grandmother? I'm building an AI conversation partner where you can practice without judgment. Looking for beta testers (30-min sessions) to shape this product! DM me.

Most people learn AI tools then hunt for problems to use them on. I learned to spot problems first, then pick the simplest tool that works. My problem-first strategy: aiweekender.substack.com/p/the-strate...

Differences between production AI systems vs. demos: Demo: Works perfectly in controlled environment Production: Handles edge cases, rate limits, and vendor outages The gap is where experience matters.

Feature engineering: Crafting mathematical transformations (log transformations, normalization, interaction terms) Prompt engineering: Crafting natural language instructions (chain-of-thought prompting, few-shot prompting etc.) Same analytical thinking, different medium. shorturl.at/ZYmJM

Old Me as Data Scientist: "We need more training data" New me as AI Engineer: "We need better examples in the prompt"

Found Google's Gemini CLI tool that writes documentation for you. Here's how: 1. Install it: npm install -g @google/gemini-cli 2. Go to your project folder: cd [path] 3. Invoke agent by running: gemini 4. Ask: "Create a comprehensive README with installation and usage examples"

ChatGPT Agent is live. It can: • Control a virtual desktop • Book travel autonomously • Build presentations We went from chatbot to digital employee in 2 years. What's the first task you'd delegate to an AI agent?

My career progression in a nutshell: Harvard taught me to analyze problems. FinTech taught me to solve them with code. Substack taught me to share the messy process.

Everyone wants to know "the secret" to breaking into data science. The secret: - Build stuff. -Post it. - Repeat. It’s just about persistence.

3 years ago, none of the AI tools you use daily existed. ChatGPT, Claude, Cursor, GitHub Copilot, LangChain - all launched in the past 36 months. The entire landscape changed faster than a college degree takes to complete.

The AI tools everyone uses today didn't exist 2 years ago. Experience matters less than you think.

Every AI application builder was a beginner 18 months ago. That's how fast this space moves.

The truth about breaking into data science: It’s 10% technical skills, 90% showing you can think through messy business problems.

Building AI is easy. Building AI that people actually want to use is hard.

Every AI demo works perfectly. AI implementations... not so much. What's been the biggest gap you've seen?

AI courses teach you knowledge. Real projects teach you judgment.

Pattern recognition brain + urban architecture = weekend sketches

Building in public is vulnerable in a way that corporate jobs aren't. When you mess up at work, your team sees it. When you mess up online, the internet sees it. But the internet also gives you further reach.

Building in public taught me the difference between perfectionism and shipping. Perfectionism keeps your work private forever. Shipping lets you learn from real users.

I'm #84 in technology on Substack! I started the AI Weekender (aiweekender.substack.com) to document my DS-to-AI transition six months ago with zero audience. I thought maybe 50 would care. This is the power of showing up consistently and sharing your learnings in public!

Every company claims they're "AI-first" now. Most are still figuring out what that means. What’s ONE effective AI use case you’ve seen actually work?

The AI job market isn't "learn Python and get hired." It's "solve problems people care about and get hired." Technical skills are table stakes. Business impact is the differentiator.

ML taught me to think in features and targets. AI taught me to think in examples and instructions.

Sometimes the answer is just a really good prompt and an API call. What’s a business problem you’ve recently solved with LLMs?

That moment when you realize your 6 years of sklearn experience doesn’t prepare you for prompt engineering… aiweekender.substack.com/p/the-ml-ski...

@ds-claudia.bsky.social shares the full breakdown of how she won $10,000 in a machine learning competition — from feature selection to threshold tuning and neural net architecture.

Traditional ML: "Let me engineer 50 features and train on 10,000 examples." Modern AI: "Let me write 3 examples in a prompt and see what happens." Read more on: The ML skills that don’t transfer to AI (and What Does) open.substack.com/pub/aiweeken...

The best API is the one you don't have to read the docs for. What's the best API you've worked with?