← Back to AI-SANA
📝 Student Learning Notes
Real learning notes from AI-SANA Certificate Program 1 students.
See how students summarized their takeaways and insights from each module in their own words.
🎬 Student-Made Video
A follow-along tutorial video made by a student. Just follow the steps shown.
▶ Watch Video
Student A
Module 1 (1.2.2) — Getting Started with Python
- First time with Python — asking Gemini questions like "oh, so that's what this code does" helped me understand.
- Adding # (hash) turns a line into a comment that the program ignores.
- Gemini can write all the code for you.
- You can paste code into Gemini and it will explain everything.
Module 2 (1.2.9) — Your First Real ML Code
- Learned what the "import" command means.
- Discovered that Gemini can't always find the right answer.
- Understanding the mechanism of training AI.
- Coding is definitely not easy.
- The reason answers were hard to find: you need to modify the example code to get the correct answer. Complete the banana/strawberry image classification example first, then solve the problems.
Module 3 (1.2.10) — Fraud Detection
- Generally, 80% of data is used for training and 20% for testing.
- max_depth=4 means the decision tree can only ask up to 4 questions.
- An F1 Score around 40% represents a well-balanced ratio.
- Understanding key terms: Accuracy, Precision, Recall, and F1 Score.
Module 4 (1.2.11) — AI in Agriculture
- The number 42 in random_state is a convention from programmers' culture, inspired by a famous sci-fi novel. Any number would work the same way.
- Threshold values 30%/50%/70%: When should you spray pesticide? Usually 70% yields the best harvest.
- There's no fixed constant for AI in agriculture — it takes experimentation and trial-and-error to find the right settings.
Module 5 (1.2.12) — Feature Engineering
- Feature Engineering: The process of transforming raw data into optimal features to maximize model performance.
- Tree depth concept — max_depth=6 means the maximum number of consecutive questions. Like a game of "20 questions."
- Decision trees find the optimal questions and order to split complex data.
- More features don't always mean better results.
- Label: The final answer (target) that the model must predict.
Module 6 (1.2.13) — Similarity & Distance
- No labels = no right answers. Unsupervised learning works by reasoning from evidence.
- Min-Max Scaling converts data to a 0–1 range to unify feature scales.
- AI only calculates similarity — the final interpretation is the human's job.
Module 7 (1.2.14) — Cosine Similarity
- Euclidean distance measures how similar data points are — the most fundamental baseline.
- Cosine similarity measures directional alignment — it cares about meaning and patterns.
- Online shopping recommendations use cosine similarity to match user preferences.
Module 8 (1.2.15) — k-Nearest Neighbors
- Predictive AI uses labeled data. K represents the number of neighbors to consult for the majority vote.
Student B
Module 1 (1.2.2) — Getting Started with Python
- First time with Python — "What is this?" Hard to understand at first.
- Asking Gemini questions — not so much "understanding" as accepting that this is how code works.
- I dropped a coding class in my education courses before — and now here I am again...
Module 2 (1.2.9) — Your First Real ML Code
- Gemini and GPT can give different (sometimes wrong) answers.
- Modifying example code to find the answer was difficult — the modification process itself felt meaningless.
- Coding is hard.
Recording your honest feelings and struggles is also an excellent learning note.
Student C
Google Colab Basics
- First time entering Google Colab.
- Google Colab = a place to build and experiment with programs together, like Docs (writing) and Sheets (spreadsheets).
- One of the most popular free tools for studying AI.
- # (comment) — anything after it is ignored by the program, but code before # will cause an error.
Image Classification
- Two fruit images: the first is labeled "0", the second "1".
- You can use images from the internet or your own computer for AI classification.
Evaluation Metrics
- Accuracy: Proportion of correct predictions. Don't rely on this alone.
- Precision: Of all fraud alerts, how many were actual fraud.
- Recall: Of all actual fraud, how many did the model catch. Currently only 10 of 33 (30.3%).
- F1 Score: Harmonic mean of precision and recall. Higher = better balance.
Module 4 (1.2.11) — Thresholds & Business Decisions
- Threshold: The cutoff point for making decisions based on model confidence. Three approaches: Paranoid / Balanced / Relaxed.
- Feature Engineering matters: Creating derived features like BMI + age risk combo yielded the highest F1 score (0.696) — better than just adding more raw features.
- Designing smart features is more important than making deeper trees.
Module 6 (1.2.13) — Unsupervised Learning & Similarity
- No labels = no answers. Unsupervised learning reasons from evidence.
- Min-Max Scaling: Normalizes data to 0–1 range to unify feature scales.
- AI only calculates similarity — final judgment is the human's responsibility.
Module 7 (1.2.14) — Distance Metrics
- Euclidean distance: Straight-line (like drone flight). Manhattan distance: Grid-based (like city walking).
- Cosine similarity: Ignores magnitude, measures only direction/pattern — core tech behind Netflix, Spotify recommendations.
Module 8 (1.2.15) — KNN Algorithm
- K=1: Extremely sensitive to noise. K=3–5: Balanced sweet spot. K=7+: Too smooth.
- K must be odd — even values cause ties in binary classification.
- There's no formula for the optimal K — you must test different values.
Module 9 (1.2.16) — ML Pipeline
- MinMax vs StandardScaler: Outliers crush MinMax scaling. StandardScaler (Z-Score) is more stable.
- Data Leakage: If test data info leaks into training, it's "cheating." Fit the scaler on training data only!
- Pipeline: Bundles preprocessing → training → prediction to prevent mistakes and leakage.
Module 10 (1.2.17) — Decision Trees & Explainability
- Black box limitation: Models that can't answer "why?" are unusable in finance, medicine, and law.
- Decision trees need no scaling and handle both numeric and categorical data.
- Overfitting dilemma: Depth 3–5 = reasonable predictions. Depth 15+ = memorizing noise → fails on new data.
Module 11 (1.2.18) — Random Forests
- Three pillars: Diversity (each tree learns different features) + Independence (no peeking) + Democracy (majority vote).
- If 94 of 100 trees say "Safe," you get 94% confidence.
- Trade-off: Random forests are black boxes but dramatically more accurate.
Module 12 (1.2.19) — Algorithmic Bias
- Night shift workers flagged 44.9% vs. 11.9% for day shift — abnormal imbalance.
- 💡 80% Rule: A minority group's pass rate must be ≥80% of the majority's. Maintenance dept at 65.6% = severe discrimination.
- "Accuracy does not equal fairness" — 90% accuracy can still hide bias.
Module 13 (1.2.20) — Regression Analysis
- Classification vs. Regression: Classification predicts categories; regression predicts continuous numbers (price, time, etc.).
- Drone flight time: Payload has 34% impact → Multiple regression achieves 83.4% R².
- Safety margin: Even if the model predicts 40-minute flight, plan for 30 minutes accounting for error.
Module 15 (1.2.22) — Hugging Face & Pre-trained Models
- 500,000+ models available on Hugging Face — dramatically saves time and cost.
- Zero-shot classification: No training data needed — just specify category names and it classifies by context.
- Sentiment analysis limitation: Overconfident 100% positive/negative even for nuanced statements.
Module 16 (1.2.23) — APIs & Business Automation
- API: The bridge that lets computers and AI exchange data automatically.
- API key security: Leaked keys can lead to massive charges. Never expose on GitHub.
- JSON + Batch processing: Automatically analyze tens of thousands of reviews in structured format.
Module 18 (1.2.25) — NLP & Prompt Engineering
- TextBlob: Free sentiment analysis library. Fast but misclassifies complex context.
- Prompt Engineering: Design AI's role (learning coach) and response format (empathy → strength → tip → closing).
- Temperature: Lower = consistent answers, higher = creative. 0.7 is a good balance.
Module 19 (1.2.26) — Reinforcement Learning & RLHF
- Reinforcement Learning: AI learns through trial-and-error with rewards and penalties.
- RLHF: Choosing your preferred chatbot response is actually training the AI's behavior.
- np.random.seed(42): Place outside the loop once for true randomness + reproducibility.
Student D
Chapter 1 — AI Coding Basics & Pre-trained Models
The goal isn't perfect understanding of Python syntax. It's about eliminating the fear of coding by running code in Google Colab with AI assistance.
- Copy, paste, run, tweak — the fundamental principle of this course.
- AI partners (ChatGPT, Claude, Gemini): Essential for explaining code and debugging.
- Key insight: Even an AI trained on hundreds of thousands of images classified a cartoon character as a "wall clock" — AI's limits are bounded by its training data.
"The era of memorizing C or Java syntax through sleepless nights is over. Coding is no longer a skill — it's the sense of flowing with AI collaboration."
Chapter 2 — Classification & Business Decision-Making
- Overall accuracy creates dangerous illusions — a model that labels everything "normal" can still score high.
- Confusion matrix analysis is essential: which error is more costly — false alarms or missed fraud?
- Threshold tuning: Connecting AI confidence to actual business costs (pesticide expense vs. crop loss) and finding the optimal break-even point.
My note: Realizing that data science is ultimately a tool for precisely controlling where company money flows — that was a revelation.
Chapter 3 — Feature Engineering & Data Transformation
- Feeding raw data to AI is like trying to use unprocessed ore as furniture.
- Creating a "high-risk combo" variable boosted F1 score significantly — data quality beats quantity.
- Data Leakage: Like secretly showing students the test answers beforehand. Extreme rigor required.
"Unorganized data is nothing but noise. That was a fresh realization."
Chapter 4 — Unsupervised Learning & Similarity
- Cosine similarity was a shock: Even if one person spends ₩1M and another ₩10K, if their spending proportions match, they're in the same taste group — math capturing the essence of human relationships.
- Cross-selling strategy: Recommend products proven popular with high-spenders to students with similar patterns.
Chapter 5 — White-box Models & Ensemble Learning
- Random Forest's principle deeply moved me — it's exactly like collective intelligence in human society.
- One stubborn executive's decision vs. diverse independent votes from many employees — the majority vote is closer to truth. AI proves this mathematically.
- For finance, medicine, and law, use interpretable models that can explain "why."
Chapter 6 — AI Ethics & Algorithmic Bias
- My belief that math-based programs are inherently fair was shattered.
- AI without malice can still become a terrible discriminator by diligently studying flawed historical practices.
- 80% Rule: Monitor regularly whether minority group pass rates reach 80% of the majority.
"How dangerous it is to hand the blade entirely to a machine — and why human oversight must persist to the end — I felt this keenly."
Chapter 7 — Regression & Uncertainty Management
- I expected AI to predict perfect future numbers. Instead, it clearly states its potential error range — refreshing shock.
- Real science isn't 100% certainty — it's precisely measuring the size of uncertainty.
- For logistics/delivery: Set safety margins at 2× the model's RMSE.
Chapter 8 — API & Workflow Automation
- What used to cost tens of millions of won and months of development now takes a few lines of API code — genuinely scary technology.
- The era of building from scratch is over. It's now about who can most cleverly combine systems by standing on giants' shoulders.
- Zero-shot classification for intelligent email routing to departments — no training data needed.
Chapter 9 — Reinforcement Learning & RLHF
- Choosing preferred chatbot responses was actually shaping the machine's personality through reinforcement learning — mind-blown.
- As a parent, training AI with rewards and penalties felt like raising a child — an uncanny emotional connection to a machine.
- Conclusion: AI is merely a probability calculator. Leaders must become orchestra conductors who precisely control thresholds and error buffers.
"That is the compass I discovered at age fifty, diving into the ocean of AI — the true compass of business that transcends eras."
Student E
Module 1 (1.2.2) — Getting Started with Python
- Copy, paste, run — you can start coding without understanding every line.
- Use AI partners (Gemini, ChatGPT, Claude) for explanations and bug fixes.
- Comments (#) make the program skip that line — useful for debugging.
- Tweak and experiment — modifying colors, sizes, speeds and building confidence through repetition.
Module 2 (1.2.9) — Machine Learning Basics
- Pre-trained models: Already trained on massive datasets, ready to use immediately.
- AI limitations & data importance: "Better data → better prediction → higher accuracy."
Module 3 (1.2.10) — Fraud Detection
- 4 features (amount, time, distance from home, new device) from bank transaction data.
- 80% training / 20% testing split.
- False Positives: Flagging legitimate customers → customer churn.
- False Negatives: Missing actual fraud → direct financial loss.
Module 4 (1.2.11) — Agriculture AI
- Beyond simple classification: "What's the exact probability this field is diseased?"
- Threshold 30% (conservative), 50% (balanced), 70% (optimistic) — each has cost trade-offs.
Module 5 (1.2.12) — Feature Engineering
- Simply adding more features → F1 score actually dropped.
- "Smart combos" win: "BMI ≥ 28 + Age ≥ 45" high-risk feature → highest F1 score.
- Decision trees work like a computer playing "20 Questions" layer by layer.
Student F — Advanced Notes
Module 1 (1.2.2) — Getting Started with Python
- Google Colab: Google's web tool for learning Python.
- Even without knowing code, ask AI to "explain like I'm an elementary school student" for easy explanations.
- Important: Watch out for Conceptually Wrong (Silent Bugs) — understanding the model's concept is crucial.
Module 5 (1.2.12) — Feature Engineering
- "Feature Engineering is more important than the algorithm itself."
- Feature Engineering is "the ability to discover hidden meaning within data."
- Key questions to ask: Can I create meaningful ratios? Differences? Interactions between variables? What derived features would a domain expert design?
Module 6 (1.2.13) — Distance Metrics
- Euclidean: Straight-line distance. Manhattan: Grid-based sum of axis distances.
- Scaling is mandatory before distance calculations.
| Scenario | Recommended Scaler |
|---|
| General ML | StandardScaler |
| KNN, K-Means | MinMaxScaler |
| Many outliers | RobustScaler |
| NLP, Sparse Data | MaxAbsScaler |
| Cosine Similarity | Normalizer |
| Heavily skewed | Log Transform |
Module 2 (1.2.9) — Classify the Strawberry
- Pre-trained models like MobileNet V2 are available — find what you need and fine-tune it.
- Always import required libraries. Understanding the model's concept prevents "conceptually wrong" (silent bugs).
Module 4 (1.2.11) — Crop Disease Detection with Thresholds
- Compare thresholds at 0.30, 0.50, 0.80 to balance False Positives vs. False Negatives for optimal farm decisions.
- Drones can collect greenness, soil humidity, plant height, and rainfall data — amazing!
- Must understand domain logic: "Rust disease thrives in humid, weakened conditions."
- Why 42? From "Hitchhiker's Guide to the Galaxy" — the answer to life, the universe, and everything. Programmer humor! 😂
Module 8 (1.2.15) — k-NN for NomadMart Loyalty
- kNN applications: Healthcare, Fraud Detection, Recommender Systems, Computer Vision.
- Different k values (1, 3, 5, 7) produce different predictions and decision boundaries.
Module 9 (1.2.16) — Feature Scaling & ML Pipelines
- MinMaxScaler is vulnerable to extreme outliers (like VIP customers). StandardScaler is more stable.
- Pipeline: Bundles all steps together. The biggest reason to use it: preventing mistakes.
Module 10 (1.2.17) — Decision Trees & Overfitting
- Overfitting: The model didn't "learn well" — it memorized the training data.
- max_depth=None → tree keeps growing until it practically memorizes each individual.
- Prevention: Limit depth (max_depth=3), set min_samples_split and min_samples_leaf.
Module 12 (1.2.19) — Algorithmic Bias
- Beyond accuracy: "Who could this model disadvantage?" is the critical question.
- Suspect the data first. Choose features carefully. Humans make the final call.
"Asking 'who could this model disadvantage?' is far more important than improving accuracy. That's the core of Responsible AI."
My note: Understanding concepts and principles comes first. Code will be handled through Vibe Coding in the AI era.
Module 13 (1.2.20) — Regression: Caspian Drone Fleet
- Built a linear regression model (y = mx + b) using drone flight distance and fuel consumption data.
- Calculated fuel costs for 3 new routes and analyzed R² for business implications.
Module 17 (1.2.24) — K-Means Clustering
- K-Means with k=2, 3, 4. At k=3, clusters named "strugglers," "crammers," and "completers."
Module 18 (1.2.25) — NLP: Student Feedback Analysis
- Used TextBlob to extract sentiment scores from 12 feedback comments for positive/negative classification.
- Identified cases where both the model and human labeler misclassified — analyzed contextual ambiguity as root cause.
Module 19 (1.2.26) — RLHF: Training AI with Human Feedback
- Analyzed 5 response pairs, voted on the better answer → trained a preference model.
- Investigated cases where human judgment and model predictions diverged.
Student G — Visual Summary Notes
Module 1 (1.2.2) — Google Colab Core Concepts
📌 Key Message
- Coding is "Copy → Run → Modify" not "Understand → Execute"
- Don't try to understand everything perfectly from the start
- "Run + Modify + Repeat" is the study method
- Actively use AI tools
When Code Breaks
- ❌ Syntax errors — from unrecognized code
- ❌ Comment (#) issues — lines with # won't execute
Debugging Methods
- Delete the error line
- Remove or add #
- Fix function names (e.g., draws → draw)
- Use Ctrl+F to find problem words
Module 2 (1.2.9) — Teachable Machine & MobileNetV2
- Pre-trained model: MobileNetV2 is already trained on ImageNet. We're "borrowing" a finished AI model.
- Computers count from 0: First banana = #0, second strawberry = #1.
- Test results: Banana 82%, Strawberry 95% — high confidence.
- Limitation found: Cartoon character → classified as "wall clock." Dog in Santa outfit → confidence dropped to 39%.
"Better data, better prediction and higher accuracy."
Student H — From the Ground Up
Module 1 (1.2.2) — Getting Started with Google Colab
- You don't need to know what every line means. Just copy generated code, paste it, and confirm it runs.
- # means "comment" — if it appears before code, it causes an error.
- Ask Gemini in Colab when stuck.
"I have zero background knowledge, so I'm starting from a blank page. I'm trying to study from the most basic concepts and ask specific questions as I go. 😅"
Core ML Concepts
- Traditional programming: Data + Rules → Answer
- Machine learning: Data + Answers → Rules/Model (computer finds patterns itself)
- Limitation: Black box problem — hard to trace "why" it reached that conclusion
- Key lesson: Garbage In, Garbage Out — bad data = bad rules
Module 3 (1.2.10) — TengriPay Fraud Detection
- Flow: Generate data → Split 80/20 → Train model → Evaluate → Business cost analysis
- False Positive: Flagging normal transactions → customer inconvenience, reputation damage
- False Negative: Missing actual fraud → direct financial loss
- Translating model errors into actual financial cost is the key insight.
Module 5 (1.2.12) — Feature Engineering
- Analogy: Like trimming ingredients for a new recipe rather than using them raw.
- Giving the model a pre-calculated BMI is more informative than separate weight and height.
- Feature quality > feature quantity — more features can actually hurt performance.
Module 10 (1.2.17) — Decision Trees & Explainability
- Decision trees = a flowchart of sequential Yes/No questions.
- Root node → first question → Internal nodes → follow-ups → Leaf nodes → final decision.
- Explainability: You can trace back "why was this loan rejected?" — the biggest advantage.
- Overfitting: Model "memorized" training data → weak against new data.
Module 11 (1.2.18) — Random Forests
- Three core principles: Diversity (each tree learns different data/features) + Independence (no influence between trees) + Democracy (majority vote).
- Individual tree errors cancel out across the forest → more robust predictions.
Module 12 (1.2.19) — Algorithmic Bias Audit
- Night shift and maintenance workers flagged at abnormally high rates due to intentionally biased data.
- AI learns biased rules from biased data — innocent people unfairly flagged.
- Key lesson: "Even if AI is smart, it's useless if the data isn't smart."
- Coding isn't just a technical skill — it affects people's lives. Always consider fairness and ethics.
Module 13 (1.2.20) — Regression: Drone Battery Prediction
- Simple linear regression: A single feature explains <50% of battery life variation.
- Multiple regression: Combining temp, wind, payload, altitude, speed significantly improves explanatory power.
- Business insight: Payload has the biggest impact on battery. High wind speed drains battery heavily.