Tredence, Quantiphi, Gramener & Sigmoid: Data / AI

CREATE TABLE sales (id INT PRIMARY KEY, region VARCHAR(20), amount DECIMAL(10,2), sale_date DATE);
INSERT INTO sales VALUES (1,'North',1200.50,'2025-07-01'), (2,'South',800.00,'2025-07-02');
UPDATE sales SET amount = amount * 1.05 WHERE region = 'North';
ALTER TABLE sales ADD COLUMN channel VARCHAR(10) DEFAULT 'online';

• Know the categories (DDL vs DML vs DCL), constraints (PK, FK, UNIQUE, CHECK) and transactions (BEGIN/COMMIT/ROLLBACK).

def majority(nums):
    cand, cnt = None, 0
    for x in nums:                       # Boyer–Moore voting
        if cnt == 0: cand = x
        cnt += 1 if x == cand else -1
    return cand if nums.count(cand) > len(nums) // 2 else None

• O(n) time, O(1) space. The final check verifies the candidate when a majority isn't guaranteed.

Answer (show your reasoning):

• "The data was tabular with 40 features and an imbalanced target, so I compared logistic regression (interpretable baseline), random forest and XGBoost with stratified 5-fold cross-validation on PR-AUC."
• "XGBoost won (0.62 against 0.48 for logistic regression). The main challenge was leakage from a post-event field; removing it dropped validation AUC but fixed the gap with the test set."
• "I'd add SHAP explanations and probability calibration next."

• Business question → data → method → insight → action, with a visual (Gramener is known for data storytelling).
• Show one strong chart that changed a decision; explain the design choices (encoding, avoiding clutter).
• Quantify the outcome and mention what you'd do differently.

• KPIs: revenue, orders, average order value, growth % (month on month and year on year), margin.
• Visuals: trend line, revenue by region and category (bar), top products (table), a map for geography, slicers (date, region, channel).
• Model: a star schema (fact sales + dimensions: date, product, customer, region); DAX measures (Total Sales = SUM(...), time intelligence with SAMEPERIODLASTYEAR).
• Design: clear titles, consistent colours, drill-through to details, tuned performance (avoid too many visuals).

• SQL: set-based work close to the data (filtering, joins, aggregations, window functions) on large tables in the warehouse. Pushing computation down avoids moving data.
• Python (pandas/PySpark): complex logic, ML feature engineering, API calls, text processing, statistical modelling, plotting.
• Rule of thumb: aggregate in SQL, then model in Python. Keep the logic in one tested place.