Ola & CRED: Data Science

• Model: ŷ = Xw + b. Cost (MSE): J(w, b) = (1/2n) Σ (ŷᵢ − yᵢ)².
• Gradients: ∂J/∂w = (1/n) Xᵀ(ŷ − y), ∂J/∂b = (1/n) Σ (ŷᵢ − yᵢ).
• Update: w ← w − α ∂J/∂w; b ← b − α ∂J/∂b. Scale the features for faster convergence. The closed form is w = (XᵀX)⁻¹Xᵀy.

WITH up AS (                                   -- distinct user-product purchases
  SELECT DISTINCT p.user_id, pr.name
  FROM purchases p JOIN products pr ON pr.id = p.product_id
)
SELECT a.name AS product_1, b.name AS product_2, COUNT(*) AS together
FROM up a JOIN up b
  ON a.user_id = b.user_id AND a.name < b.name   -- each unordered pair once, alphabetical
GROUP BY a.name, b.name
ORDER BY together DESC, product_1, product_2
LIMIT 5;

• At scale: the self-join explodes for heavy users, so cap the products per user, pre-aggregate, or approximate with sampling or MinHash. State the ordering convention the question asks for (this one wants the alphabetically first item listed second, so swap the output columns if needed).

• North star: monthly active payers paying bills on time through CRED (core behaviour), plus retention.
• Supporting metrics:
• Reward redemption rate.
• Repeat payments.
• Payment value share.
• Engagement with new features.
• Referral rate.
• Economics: reward cost per incremental active payer, LTV/CAC, and the margin from partners.
• Guardrails: fraud and abuse rates, user complaints, payment failure rates.
• Causality: holdout groups for reward changes; cohort analysis.

• Idea: when demand exceeds supply in a zone and time window, raise the price multiplier to reduce demand and attract drivers, restoring wait times.
• Model: 1. Forecast demand and supply per zone for the next 15–30 minutes (gradient boosting or spatio-temporal models using events, weather, time). 2. Estimate demand elasticity to price. 3. Choose a multiplier that meets a target ETA or fulfilment rate, with caps and smoothing (no sudden jumps) and fairness constraints.
• Evaluate: completed rides, ETA, cancellations, driver earnings, rider complaints; switchback experiments.

• Trees capture non-linearities and interactions automatically, handle mixed feature types and missing values (XGBoost), and need no scaling.
• Logistic regression is more interpretable and well calibrated, and better with sparse high-dimensional text or very little data.
• Choice: cross-validated performance plus the need for explainability. Use SHAP to explain the tree models.

• Data access delays: "I built the pipeline against a sample schema while approvals were pending, so I was ready on day 1 of access."
• Unclear metric definitions: "I aligned with the PM on a written definition doc."
• Model–business mismatch: "High AUC, but ops couldn't act on 30% of the flags. I retuned the threshold to the ops team's capacity."
• Each hurdle should show proactiveness and measurable impact.