In my previous post (Intro to ML #9), I covered ways to get more out of generative AI — prompts, RAG, MCP, and agents — based on my MBA data science course.
This is the wrap-up. The final session of the course was “Demo Day” — a chance to use everything we’d learned across the previous five sessions to define our own problem, build something, and present it to the class.
In this post I’ll share what I built on Demo Day, what I took away from my classmates, and how I’m organizing the course’s lessons through a What / So what / Now what frame.
What I Built: An ISO Audit Self-Check System
I picked the generative AI track. My theme: “Make audits less painful for everyone.” I work on ISO-related audits in my day job, and I wanted to address a frustration I’d carried for years.
The problem I wanted to solve
- Audit standards and checkpoints are voluminous and hard for the field to navigate
- Different reviewers reach different conclusions on the same scene (high subjectivity)
- Manually mapping photos to clauses is tedious every single time
The system I built
Using Dify combined with OpenAI, I built a workflow that takes ISO standards (text) and on-site photos (data centers, offices) as inputs, then returns a self-check result mapped to the relevant audit points. I loaded the standards into Dify’s knowledge base, then passed the photos to a vision-capable LLM with structured prompts that produced a judgment.
Where I struggled
Three walls.
- Inconsistent judgments: The same photo would yield different verdicts on different runs. The “butterfly effect of prompts” I wrote about in #9 was very real here. I had to spell out the evaluation criteria, the output format, and the granularity of judgment in the prompt itself before things stabilized.
- Latency growing with complexity: Each generative AI step added meaningful response time. Chaining RAG and LLM calls quickly compounded into unacceptable wait times. I ended up trimming steps and reserving generative AI only for the parts that genuinely needed it.
- Linking standards to physical scenes: To say “this part of the photo doesn’t comply,” I needed clean correspondence between specific clauses and visual elements. I had to make “always cite the relevant clause number” an explicit prompt rule.
When the system finally worked end to end, the pieces I’d been touching across the course — problem framing, data prep, model selection, evaluation, prompt design — all clicked into one coherent thing. That felt different from understanding them individually.
What I Took Away from Classmates’ Projects
To be honest, my project wasn’t selected for any of the recognition spots. Watching the projects that were celebrated, I realized something: the way they identified problems, and the way they pictured the people they were solving for, were on a different level.
Without naming specific projects, what struck me was how many of them started from small inconveniences in everyday life — family situations, parenting moments, the little decisions that quietly drain a day. The shape of the question was: “What if I used AI to make the next few meters of my own life easier?” I’d anchored on workplace problems, which kept me in an “efficiency optimization” frame.
The other thing was delight. Even themes that made me think “why would you spend time on that?” became compelling when the creator was visibly enjoying themselves. AI is a tool for solving problems, but it’s also a tool for giving form to curiosity. I’d lost sight of the second part.
Course Reflection: What / So what / Now what
What? — The core insight
Both machine learning and generative AI are, at the bottom of it all, doing prediction. The first predicts the next number or label from data; the second predicts the next word or pixel from context. The base mechanism is the same — the inputs and the shape of the output differ. Internalizing that one observation raised my comprehension of AI news and papers by a clear notch.
The other one was the Issue × Data × Algorithm framing. Strengthening one without the others doesn’t generate value. You need a worthwhile question, data that matches it, and an algorithm that can act on both. Demo Day was the visceral version of that lesson.
So what? — Implications for work
The era of debating “can AI be used here?” is over. The real question has shifted to “where in your work, and how, does it get embedded?” And to embed it well, you have to articulate what you’re trying to solve in your own words. That’s not a technical skill — it’s a business framing skill.
The other lesson is that you don’t understand it until you’ve touched it. Running a prediction model in Exploratory or assembling an agent in Dify made me feel things — the messiness of real data, the fragility of prompts — that no amount of reading would have produced.
Now what? — What I’m actually doing next
Three commitments.
- Ship the audit system at work: I’m not letting the Demo Day version be the end. I’ll iterate it into something usable in the actual audit process. If it scales internally, it could meaningfully reduce the subjectivity problem.
- Adopt the “next few meters” frame: I’ll deliberately look for AI use cases in my own life and household, not just at work. That’s the lesson I’m taking from my classmates.
- Touch it daily: A line from the course stuck with me — “this is a marathon, not a sprint.” I’ll use AI somewhere every day. The course confirmed for me that the gap between people who use it daily and those who don’t widens fast.
Wrapping Up the Series
This is the tenth post in the series. Writing my way through it deepened the material in a way that felt similar to the hands-on work in the course itself. The distance between “I get it” and “I can explain it” turned out to be much larger than I’d assumed.
The next post covers the other major topic from the final session — AI and intellectual property — written for international readers who do business in or with Japan. I’ll explain why Japan is sometimes called a “machine learning paradise” and what the practical implications are for how you handle data, models, and AI-generated content.
→ Intro to ML #11 — AI and Copyright in Japan
Books That Helped Me Reflect
① For seeing AI as a personal journey
Fei-Fei Li’s The Worlds I See is part memoir, part history of computer vision, part argument for human-centered AI. Reading it alongside an MBA data science course was a useful counterweight to the algorithm-and-business framing I was getting in class. It gave the field a face.
② For making the case to keep being a generalist
David Epstein’s Range argues that in a world of specialists, generalists who sample broadly tend to win the long game. As someone studying machine learning while running an IT delivery practice, I needed this book’s permission to keep being a generalist. It pairs well with the “why am I learning this at my age?” feeling that any mid-career learner will recognize.