Swiggy: Data Science (Campus 2025)

• Data structure: a trie of normalised queries, where each node stores the top-k completions by popularity. Lookup is O(prefix length).
• Ranking signals: global popularity, location (only serviceable restaurants nearby), time of day, the user's history, open/closed status, freshness of trends.
• Robustness: typo tolerance (edit distance or n-gram index), transliteration ("paneer" vs "panir"), synonyms ("biryani" ≈ "biriyani").
• Serving: precomputed top-k per prefix per city in Redis; a latency budget under 50 ms; update the counts in near real time from logs.
• Evaluation: keystrokes saved, CTR on suggestions, search-to-order conversion; A/B test ranking changes.

• L2 (Ridge): penalty λΣw². Shrinks weights smoothly and handles correlated features. The usual default.
• L1 (Lasso): penalty λΣ|w|. Produces sparse weights (feature selection), good with many irrelevant features.
• Elastic Net mixes both. In neural networks, weight decay (L2) plus dropout.
• Tune λ by cross-validation and standardise features first.

• Decompose: preparation time + rider assignment wait + rider to restaurant + restaurant to customer.
• Features:
• Restaurant prep history by dish and time of day.
• Current kitchen load.
• Rider availability.
• Distance and road-graph ETA.
• Traffic, weather, rain flags.
• Festival days.
• Model: gradient boosting per component, or one model with quantile loss. Predict P50 and P90 so you can promise conservatively.
• Metrics: MAE, plus the share of orders delivered later than promised (customer trust). Monitor by city and zone; retrain often.

WITH c AS (
  SELECT o.customer_id, r.cuisine, COUNT(*) AS n,
         RANK() OVER (PARTITION BY o.customer_id ORDER BY COUNT(*) DESC) AS rk
  FROM orders o JOIN restaurants r ON r.restaurant_id = o.restaurant_id
  GROUP BY o.customer_id, r.cuisine
)
SELECT customer_id, cuisine, n FROM c WHERE rk = 1;

• A window function over an aggregate works in a single CTE. RANK keeps ties; use ROW_NUMBER to pick exactly one.

• Signals: device and payment reuse across accounts, new accounts that redeem immediately, mismatches between location and restaurant, repeated maximum-discount bills, reviews posted only by linked accounts.
• Methods: velocity rules, a graph of shared identifiers (connected components), an anomaly score, then a supervised model once labels exist.
• Action: limit offers for risky clusters and require verification. Measure the false-positive impact on genuine users.

from collections import Counter
def top_dishes(orders, k=3):
    return Counter(d for o in orders for d in o["items"]).most_common(k)

• O(n) to count plus O(m log k) for the top k. Mention streaming with a heap if the data is huge.