JPMorgan & Amex: AI / Data Science (2025)

import pandas as pd, numpy as np, matplotlib.pyplot as plt

df = pd.read_csv("transactions.csv")
print(df.shape, df.dtypes, df.isna().mean().round(3), sep="\n")
print(df.describe(percentiles=[.01, .25, .5, .75, .99]).T)

col = "amount"
q1, q3 = df[col].quantile([.25, .75]); iqr = q3 - q1
lo, hi = q1 - 1.5 * iqr, q3 + 1.5 * iqr
out = (df[col] < lo) | (df[col] > hi)
print(f"outliers: {out.mean():.2%}")

fig, ax = plt.subplots(1, 2, figsize=(10, 4))
ax[0].boxplot(df[col].dropna()); ax[0].set_title("Amount (raw)")
ax[1].hist(np.log1p(df[col].clip(lower=0)), bins=50); ax[1].set_title("log(1+amount)")
plt.tight_layout(); plt.show()

df["amount_capped"] = df[col].clip(lo, df[col].quantile(.995))   # winsorise, keep the rows
df["is_outlier"] = out.astype(int)                              # keep the signal for fraud models

• Say why: in banking, outliers are often the signal (fraud), so flag them rather than deleting them. Cap them for linear models, and log-transform skewed money amounts.
• Also mention z-scores vs IQR (IQR is robust to skew), and Isolation Forest for multivariate outliers.

• First ask whether the outlier is an error or a real rare event. Data-entry errors get fixed or removed; real extremes (large transactions) are kept.
• Options:
• Winsorise or cap.
• Log or Box-Cox transform.
• Robust scaling (median/IQR).
• Tree models, which are insensitive to outliers.
• Robust losses (Huber).
• A separate model for the extreme segment.
• Multivariate outliers: Isolation Forest, LOF, Mahalanobis distance.
• Document the decision and its effect on model metrics; regulators may ask.

• Real-time scoring within about 100 ms at authorisation.
• Features:
• Velocity (transactions in 1 h / 24 h).
• Amount vs the user's typical spend.
• Merchant risk, geography and travel distance.
• Device and channel.
• Time of day.
• Graph links (shared cards, devices, merchants).
• Model: gradient boosting for tabular data plus rules for known patterns; consider a sequence model for transaction history.
• Imbalance: class weights; PR-AUC; recall at a fixed false-positive rate.
• Decision layer: approve / step-up (OTP) / decline, with thresholds tuned on fraud loss vs customer friction.
• Operations:
• Labels arrive with a delay (chargebacks), so handle that.
• Monitor drift, retrain weekly or monthly.
• Model risk management (SR 11-7 style validation) and explainability via reason codes.

• Common use cases:
• Internal knowledge assistants (policies, procedures).
• Document extraction (KYC, trade documents).
• Call summarisation.
• Code assistants.
• Research summarisation for analysts.
• Extra constraints:
• Data residency.
• No customer PII leaving the bank's boundary (private endpoints or self-hosted models).
• Explainability and audit trails.
• Model risk governance.
• Strict accuracy on numbers.
• Patterns: RAG with citations, a human in the loop for any customer-facing or decision output, deterministic tools for calculations (never let the LLM do the arithmetic), and red-teaming for prompt injection.

Answer (show the structure, not the exact number):

• Assume about 100 million active cards, with about 1.5 transactions per card per day → about 150 million transactions per day.
• 150M / 86,400 s ≈ 1,700 TPS on average. Peak (festive sales, evenings) is roughly 3–5× → design for about 8,000 TPS.
• At a p99 latency budget of about 50 ms per score, with each instance handling about 500 requests per second → about 16 instances, plus headroom across 2 regions.
• State your assumptions clearly, sanity-check them, and tie them to the design (caching features in Redis, horizontal scaling).

• The problem: in backpropagation through many layers, gradients are products of derivatives. With sigmoid or tanh (derivatives below 1), they shrink exponentially, so early layers barely learn. RNNs suffer this over time steps.
• Fixes:
• ReLU / GELU activations.
• Residual connections (ResNets, Transformers).
• Normalisation (Batch / Layer Norm).
• Careful initialisation (He / Xavier).
• LSTM / GRU gates for sequences.
• Gradient clipping, for exploding gradients.

• Own it: "Our credit model failed independent validation on stability (PSI 0.3 in one segment)."
• Act: 1. Run a root-cause review with the team, without blame. 2. Find the issue: a feature driven by a policy change. 3. Rebuild with stable features and add segment-level monitoring. 4. Communicate the timeline to the business and risk teams every week.
• Result: "It passed revalidation in five weeks. We added stability tests to our standard development checklist, and the next two models passed on the first submission."
• A VP-level answer shows governance, communication, coaching and process improvement.