1Pick a question
Choose one, or let Pondly pick. Set the difficulty; it stays the same for the whole question.
For AI, machine learning and data science interviewsCode is written by AI now.
Code is written by AI now.
Interviews now test your judgment.
Land your dream data job. Practice with an AI interviewer built for data scientists, ML and AI engineers, and learn the part AI can’t do for you.
- Class imbalanceA model hits 99% accuracy on fraud detection. Is that good?
- Data drift vs concept driftYou ship a churn model and precision drops from 0.48 to 0.31 after two weeks. What do you check first?
- Feature importanceA stakeholder asks why the model declined a particular customer. The model is gradient boosting. What do you tell them?
- A/B testingA new feature shows an 8% lift in the first week of an A/B test, but the gap shrinks to nothing by week three. What's going on?
- Debug a broken RAG systemYour RAG system suddenly starts giving incorrect answers. What's the first thing you investigate, and how would you prove that's actually the root cause?
- Evaluating LLM outputYou changed a prompt in your LLM pipeline. How do you know whether it actually made things better?
Why it mattersAnyone can generate the code.
Anyone can generate the code.
The interview is about defending it.
Track What the AI writes in seconds What the interviewer actually asks
01Concepts
X_train, X_test = train_test_split(
df, test_size=0.2)02System design
scores = model.predict(user_item_pairs)
top_k = scores.argsort()[-10:]03Troubleshooting
df["income"] = df["income"].fillna(
df["income"].median())04Product sense
dau = events.groupby("date")["user_id"].nunique()
dau.pct_change().tail(7)Inside a Pondly session
It follows your reasoning. Then it pushes back.
Interviewer
Your RAG system suddenly starts giving incorrect answers. What’s the first thing you investigate, and how would you prove that’s actually the root cause?
Me
I’d split retrieval from generation first. Pull a sample of the bad answers and check whether the right documents were in the top results. If they weren’t, it’s retrieval, not the model.
Interviewer
Separating retrieval from generation is a strong first move: it halves where the fault can be. Say the right chunks are coming back. What changed recently, and how would you show it’s the cause rather than a coincidence?
Topics
Every topic they’ll push you on.
- Class imbalance
- Bias-variance trade-off
- Confidence intervals
- A/B testing
- p-values
- Simpson's paradox
- Multiple testing correction
- Bootstrapping
- Calibration
- Precision and recall
- ROC AUC
- Cross-validation
- Hypothesis testing
- Correlation vs causation
- Uplift modelling
- Difference-in-differences
- RAG
- Vector databases
- Hallucination
- Evaluating LLM output
- Prompt injection and guardrails
- Fine-tuning vs RAG
- Chunking strategies
- Reranking
- Embeddings
- Transformers
- Agents
- Temperature and sampling
- Tokens and the context window
- Structured output and function calling
- Vector search vs keyword search vs hybrid
- Debug a broken RAG system
- Data drift vs concept drift
- Data leakage
- Train/serve skew
- Feature importance
- Gradient boosting
- Missing values
- Batch vs online prediction
- Deployment strategies
- Retraining triggers
- Choosing a classification threshold
- Regularisation
- Ridge vs Lasso vs elastic net
- Multicollinearity
- Feature selection
- Overfitting and underfitting
- Baseline models
How it works
From question to debrief in three steps.
2Talk it through
The interviewer asks, listens and follows up. Stuck? Hints build from a few ideas to the full answer.
3Read the debrief
Four ratings with evidence from your own words, one thing to keep, one habit to sharpen.