Interview Preparation System

Definition: A sequential manufacturing arrangement where a product moves through a series of workstations, each performing a specific operation, until the product is complete.

Why it matters: The fundamental structure of mass production — understanding flow, station sequencing, and balance directly impacts throughput.

💬 "Walk me through how an assembly line is structured and what makes it efficient or inefficient."

Definition: The total time from the beginning to the end of a process — from when you start working on one unit to when you start on the next.

Why it matters: Directly determines your production rate. Reducing cycle time increases throughput.

💬 "What is the difference between cycle time and takt time, and why does it matter for line balancing?"

Definition: The rate at which products must be completed to meet customer demand. It is the "heartbeat" of production.

Why it matters: Takt time sets the pace of the entire line. If cycle time > takt time, there is a bottleneck and customer demand will not be met.

💬 "If available time is 480 min/day and demand is 240 units/day, what is the takt time and what does it tell you?"

Definition: The rate at which a system produces output — units per hour, per day, or per shift.

Why it matters: The ultimate measure of a production system's output capability. Constrained by the bottleneck (Theory of Constraints).

💬 "How would you identify and increase throughput in a constrained system?"

Definition: The workstation or process step with the lowest capacity, which limits the overall throughput of the entire system.

Why it matters: An hour lost at the bottleneck is an hour lost for the entire system. All improvement effort should focus on the bottleneck first.

💬 "How would you identify the bottleneck in a 60-station assembly line using data?"

Definition: A composite measure of manufacturing productivity that combines Availability, Performance, and Quality into a single score.

Why it matters: World-class OEE is ~85%. It identifies hidden capacity losses.

Availability = Run Time / Planned Production Time | Performance = (Ideal Cycle Time × Units) / Run Time | Quality = Good Units / Total Units

💬 "If OEE is 62%, which component is most likely the culprit and how do you investigate?"

Downtime: Time when equipment or a workstation is not producing. Rework: Units that fail quality and must be reprocessed. Scrap: Units that fail quality and cannot be recovered.

Why it matters: All three represent pure waste — lost time, material, and labour with zero customer value added.

💬 "How do you distinguish between planned and unplanned downtime, and why does this matter?"

Definition: The process of distributing work tasks evenly across workstations so that cycle time at each station is as close to takt time as possible, minimising idle time.

Why it matters: Imbalanced lines create bottlenecks and idle stations simultaneously, wasting both labour and time.

💬 "In a 60-station line, how would you use data to identify which stations are overloaded?"

Definition: Inventory that has started production but has not yet been completed. Units sitting between workstations.

Why it matters: High WIP indicates an imbalanced line or bottleneck. Little's Law: WIP = Throughput × Cycle Time.

💬 "If WIP suddenly increased between two specific stations, what would you investigate first?"

Defect Rate: The proportion of units produced that are defective. First Pass Yield (FPY): The percentage of units that pass quality inspection on the first attempt without rework.

💬 "What is the difference between defect rate and FPY, and which is a better quality KPI?"

Definition: A statistical technique that identifies the vital few causes responsible for the majority (typically 80%) of defects or problems.

Why it matters: Focuses improvement effort where it has the highest impact. In defect analysis, finding the top 3 defect types that cause 80% of failures directs your root cause investigation.

💬 "In your Tata Motors project, how did you prioritise which defects to investigate first?"

5 Why: A root cause analysis technique where you iteratively ask "Why?" five times (or until the root cause is found). Fishbone: A structured diagram that maps potential causes of a problem across categories: Machine, Man, Method, Material, Measurement, Environment (6M).

Why it matters: Surface symptoms are never the root cause. These tools ensure you fix the right problem.

💬 "Walk me through applying 5 Why to a defect type you found in your dataset."

Definition: SPC uses statistical methods to monitor and control a manufacturing process. Control charts plot process measurements over time with Upper and Lower Control Limits (UCL/LCL) calculated at ±3σ from the mean.

Why it matters: Distinguishes common cause variation (normal process noise) from special cause variation (a real problem requiring investigation).

💬 "What does a point outside the control limits tell you about a manufacturing process?"

Cp: Measures how spread the process is relative to the specification width — does the process fit within the spec limits? Cpk: Measures whether the process is centred within spec limits. Cpk ≥ 1.33 is generally acceptable.

💬 "If Cp = 1.5 but Cpk = 0.8, what does this tell you about the process?"

PPM (Parts Per Million): Number of defective units per million units produced. DPMO (Defects Per Million Opportunities): Number of defects per million opportunities for a defect to occur — accounts for units with multiple quality check points.

💬 "Six Sigma means 3.4 DPMO. Why does DPMO matter more than defect rate in high-volume manufacturing?"

Definition: A proactive risk assessment technique that identifies potential failure modes in a process, their effects, and their causes — before failures occur. Scored using RPN = Severity × Occurrence × Detection.

Why it matters: Prioritises quality prevention efforts based on risk.

💬 "How is FMEA used in a manufacturing environment and what does a high RPN indicate?"

Definition: A workplace organisation system: Sort (Seiri), Set in Order (Seiton), Shine (Seiso), Standardise (Seiketsu), Sustain (Shitsuke).

Why it matters: Reduces waste from searching, motion, and accidents. Creates the discipline foundation for all other lean tools.

💬 "If you were implementing 5S in a 60-station assembly line, where would you start?"

Definition: The 8 categories of waste in lean manufacturing — Transport, Inventory, Motion, Waiting, Overproduction, Over-processing, Defects, Skills (unused talent).

Why it matters: Mnemonic TIMWOODS. In your project: Waiting (machine downtime), Defects, Over-processing, and Skills (unutilised cross-skilling) were likely all present.

💬 "Which of the 8 wastes would most directly link to operator skill levels?"

Kaizen: Japanese for "change for better" — a philosophy of continuous, incremental improvement involving all employees at all levels.

Why it matters: Sustains gains and builds an improvement culture. Contrasts with kaikaku (radical change/innovation).

💬 "How is Kaizen different from a major process redesign and when do you use each?"

JIT (Just-in-Time): A production philosophy where materials and parts arrive exactly when needed — not before, not after — eliminating inventory waste. Kanban: A visual pull-signal system that triggers replenishment only when downstream demand occurs.

Why it matters: JIT minimises inventory cost and waste; Kanban is the execution mechanism. Toyota Production System is built on these.

💬 "What are the risks of JIT, especially for a company like Tata Motors?"

Definition: A mechanism or process design that prevents errors from occurring or makes errors immediately obvious when they do occur.

Why it matters: Shifts quality control from inspection (detecting defects after) to prevention (stopping defects at source).

💬 "Give an example of a poka-yoke in a manufacturing context and how it relates to reducing defect rates."

Definition: A lean tool that visually maps all steps (value-adding and non-value-adding) in the flow of a product from raw material to customer — identifying waste and improvement opportunities.

Why it matters: Often reveals that <5% of total lead time is actual value-added time. The rest is waiting, movement, and inspection.

💬 "How would you use VSM to identify where skill gaps cause the most disruption in a line?"

Definition: A structured grid that maps each employee/operator against each operation or skill, with their proficiency level (e.g., 0=Untrained, 1=Aware, 2=Partially Skilled, 3=Skilled, 4=Expert/Can Teach).

Why it matters: Enables identification of skill gaps, cross-training needs, single points of failure, and correlation between skill and quality outcomes.

💬 "How did you construct the skill matrix in your Tata Motors project and what rating scale did you use?"

Definition: Analytical approach linking operator skill ratings to the defect rate at their assigned station, to determine whether skill is a statistically significant predictor of defect occurrence.

Why it matters: This is the core hypothesis of your Tata Motors project. If skill level predicts defect rate, targeted training becomes a quality intervention.

💬 "What was your dependent variable, what were your independent variables, and what statistical test did you use to test this relationship?"

Cross-Skilling: Training employees to perform tasks in adjacent roles or stations, increasing workforce flexibility. Multi-Skilling: Training employees to perform multiple distinct skills across different departments or functions.

Why it matters: Reduces absenteeism-driven bottlenecks, improves line rebalancing flexibility, and reduces defects caused by single-skilled operators covering unfamiliar stations.

💬 "How would you use your skill matrix to design a cross-skilling program?"

Definition: The process of assigning the right operator (with the right skill level) to the right station to optimise quality and throughput outcomes.

Why it matters: Optimal allocation means high-skill operators at defect-prone stations, not arbitrary or seniority-based assignment.

💬 "How would you recommend workforce allocation decisions based on your skill matrix and defect data?"

Job Shop: Custom, one-off products. High variety, low volume. (e.g., custom machinery) | Batch: Groups of identical products made together. (e.g., biscuits, pharmaceuticals) | Mass: High-volume standardised products on assembly lines. (e.g., automobiles) | Continuous: Uninterrupted 24/7 production of fluid products. (e.g., oil refining, cement)

Why it matters: Each system has a different cost structure, flexibility, and quality challenge. Tata Motors uses mass production.

💬 "What are the trade-offs between job shop and mass production in terms of flexibility, quality, and cost?"

Design Capacity: Maximum output under ideal conditions. Effective Capacity: Maximum output accounting for planned losses (maintenance, breaks). Utilisation: Actual output as % of design capacity.

Theory of Constraints (TOC): Goldratt's framework — every system has one constraint (bottleneck) that limits throughput. The 5 steps: Identify → Exploit → Subordinate → Elevate → Repeat.

💬 "If a plant is running at 65% utilisation, is that a problem? What other data would you need?"

Aggregate Planning: Medium-term (3–18 months) production planning that determines workforce levels, inventory, and production rates to meet forecasted demand at minimum cost.

S&OP (Sales & Operations Planning): A cross-functional monthly process aligning Sales, Finance, and Operations on a single integrated plan.

Why it matters: Bridges long-term strategy and short-term scheduling. Prevents costly reactive decisions.

💬 "What is the difference between chase strategy and level strategy in aggregate planning?"

Definition: The optimal order quantity that minimises total inventory costs — the trade-off between ordering cost and holding cost.

Where D = Annual demand, S = Ordering cost per order, H = Annual holding cost per unit.

Why it matters: Ordering too often increases ordering costs; ordering too much increases holding costs. EOQ finds the sweet spot.

💬 "What assumptions does EOQ make and when do those assumptions break down?"

ROP (Reorder Point): The inventory level at which a replenishment order should be placed to avoid stockout during lead time.

Z = service level factor (e.g., Z=1.65 for 95%, Z=2.33 for 99%), σ = standard deviation of demand.

Why it matters: Higher service level → more safety stock → higher holding cost. This is the core operational trade-off.

💬 "How does demand variability affect safety stock requirements? What if lead time is also variable?"

ABC: Classifies inventory by value — A items (top 10% of SKUs, ~70% of value), B items (20%, ~20% of value), C items (70% of SKUs, ~10% of value).

VED: Classifies by criticality — Vital (operation stops without it), Essential, Desirable. FSN: Classifies by movement — Fast, Slow, Non-moving.

Why it matters: Not all inventory deserves the same management attention. A-VED items get tightest control.

💬 "In a manufacturing context, why might a C-category item (by value) still be vital to manage carefully?"

Inventory Turnover: How many times inventory is sold and replaced in a period. Higher is generally better (less working capital locked up).

Why it matters: Low turnover = cash tied up in inventory. High turnover = lean but risks stockouts.

💬 "If a company's inventory turnover drops from 8x to 5x, what are the most likely causes?"

MRP (Material Requirements Planning): A system that determines what materials are needed, in what quantities, and when — based on production schedules and Bills of Materials.

Bullwhip Effect: Small variations in consumer demand amplify as you move upstream in the supply chain — leading to large swings in orders for suppliers.

Supply Chain Resilience: The ability of a supply chain to anticipate, prepare for, respond to, and adapt to disruptions.

💬 "What causes the bullwhip effect and what strategies can operations managers use to dampen it?"

INNER JOIN: Returns rows where there is a match in both tables. LEFT JOIN: Returns all rows from the left table, matched rows from right (NULLs where no match). FULL OUTER JOIN: Returns all rows from both tables. SELF JOIN: Joins a table to itself (e.g., find employees with the same manager).

-- Find operators who have defects but join with their skill ratings
SELECT o.operator_id, o.name, s.skill_level, COUNT(d.defect_id) as defect_count
FROM operators o
LEFT JOIN skill_matrix s ON o.operator_id = s.operator_id AND s.operation_id = d.operation_id
LEFT JOIN defects d ON o.operator_id = d.operator_id
GROUP BY o.operator_id, o.name, s.skill_level
ORDER BY defect_count DESC;

💬 "What is the difference between a LEFT JOIN and INNER JOIN, and when would you get different results?"

Window functions perform calculations across a set of rows related to the current row — without collapsing rows like GROUP BY does.

-- Rank stations by defect count, partition by shift
SELECT station_id, shift, defect_count,
  RANK() OVER (PARTITION BY shift ORDER BY defect_count DESC) as rank_in_shift,
  LAG(defect_count) OVER (PARTITION BY station_id ORDER BY date) as prev_day_defects
FROM production_data;

💬 "How do you calculate a 7-day rolling average defect rate per station using SQL?"

-- Q1: Top 3 defect types per station
SELECT * FROM (
  SELECT station_id, defect_type, count(*) as freq,
    RANK() OVER (PARTITION BY station_id ORDER BY COUNT(*) DESC) as r
  FROM defects GROUP BY station_id, defect_type
) t WHERE r <= 3;

-- Q2: Month-over-month defect rate change
SELECT month,
  defect_rate,
  LAG(defect_rate) OVER (ORDER BY month) as prev_month,
  defect_rate - LAG(defect_rate) OVER (ORDER BY month) as mom_change
FROM monthly_quality;

-- Q3: Operators with skill < 2 but high defect rates
SELECT o.name, s.avg_skill, d.defect_rate
FROM operators o
JOIN (SELECT operator_id, AVG(skill_level) as avg_skill FROM skill_matrix GROUP BY operator_id) s ON o.id=s.operator_id
JOIN (SELECT operator_id, SUM(defects)*1.0/SUM(units) as defect_rate FROM production GROUP BY operator_id) d ON o.id=d.operator_id
WHERE s.avg_skill < 2 AND d.defect_rate > 0.05
ORDER BY d.defect_rate DESC;

💬 "Write a query to find stations where defects increased over the last 3 months consecutively."

Key techniques for cleaning large manufacturing datasets:

import pandas as pd
import numpy as np

# Load
df = pd.read_csv('production_data.csv', low_memory=False)
print(df.shape, df.dtypes)

# Missing values
print(df.isnull().sum())
df['skill_level'].fillna(df.groupby('station_id')['skill_level'].transform('median'), inplace=True)

# Duplicates
df.drop_duplicates(subset=['record_id', 'date', 'operator_id'], keep='first', inplace=True)

# Outliers (IQR method)
Q1 = df['cycle_time'].quantile(0.25)
Q3 = df['cycle_time'].quantile(0.75)
IQR = Q3 - Q1
df = df[(df['cycle_time'] >= Q1-1.5*IQR) & (df['cycle_time'] <= Q3+1.5*IQR)]

# Data types
df['date'] = pd.to_datetime(df['date'])
df['shift'] = df['shift'].astype('category')

# Defect rate per station
defect_rate = df.groupby('station_id').apply(
    lambda x: x['defect'].sum() / len(x)
).reset_index(name='defect_rate')

💬 "When you say you cleaned ~1 million records — what specific steps did that involve?"

from scipy import stats
import statsmodels.api as sm

# Correlation: skill level vs defect rate
corr, p_value = stats.pearsonr(df['avg_skill'], df['defect_rate'])
print(f"Correlation: {corr:.3f}, p-value: {p_value:.4f}")

# Spearman (non-parametric alternative if not normally distributed)
spearman_corr, sp_pval = stats.spearmanr(df['avg_skill'], df['defect_rate'])

# ANOVA: defect rate across skill level groups
groups = [df[df['skill_group']==g]['defect_rate'] for g in df['skill_group'].unique()]
f_stat, anova_p = stats.f_oneway(*groups)

# Linear Regression
X = sm.add_constant(df['avg_skill'])
model = sm.OLS(df['defect_rate'], X).fit()
print(model.summary())

💬 "How did you validate that your correlation was statistically significant and not due to chance?"

H₀ (Null): No relationship / no difference / no effect. H₁ (Alternative): There is a relationship / difference / effect.

p-value: Probability of observing results at least as extreme as yours, assuming H₀ is true. If p < α (typically 0.05), reject H₀.

Type I Error: Rejecting H₀ when it is actually true (false positive). Type II Error: Failing to reject H₀ when it is actually false (false negative).

💬 "You found p=0.03 for the skill-defect correlation. What does this mean and what doesn't it mean?"

Linear Regression: Models the relationship between a continuous dependent variable and one or more independent variables. R²: Proportion of variance in the dependent variable explained by the model (0–1). Logistic Regression: When dependent variable is binary (defect: yes/no).

Assumptions: Linearity, Independence, Homoscedasticity, Normality of residuals (LINE).

💬 "In your project, why might logistic regression be more appropriate than linear regression?"

Correlation: Two variables move together statistically. Does NOT imply one causes the other. Causation: One variable directly produces a change in another.

Why it matters: In your project, you can show correlation between skill level and defects. To establish causation you would need a controlled experiment (e.g., randomised training intervention and measuring defect change).

Confounders: Hidden variables that affect both (e.g., station complexity affects both required skill AND defect rate — your skill-defect correlation might be driven by station complexity, not skill directly).

💬 "How would you establish that low skill level CAUSES higher defects, not just correlates with them?"

Trend: Long-term increase or decrease in data. Seasonality: Repeating pattern at fixed calendar intervals (weekly, monthly, annual). Cyclicality: Irregular up-down movements over multi-year business cycles (not fixed frequency). Stationarity: Statistical properties (mean, variance) do not change over time — required for ARIMA.

Autocorrelation (ACF): Correlation of a time series with its own past values. PACF: Partial autocorrelation after removing effects of intermediate lags. Used to identify ARIMA p and q parameters.

💬 "How do you test for stationarity and what do you do if the series is non-stationary?"

Naive: Forecast = last period's actual. Simple baseline. | Moving Average: Average of last n periods. Smooths noise. | Exponential Smoothing (SES): Weighted average — more recent periods weighted more (α parameter). | Holt: SES + trend component. | Holt-Winters: SES + trend + seasonality (additive or multiplicative).

Croston's Method: Designed for intermittent demand — separately models the interval between demand occurrences and the demand size when it occurs. Avoids over-forecasting zeros.

ARIMA(p,d,q): AutoRegressive Integrated Moving Average. p=AR order (past values), d=differencing (to achieve stationarity), q=MA order (past forecast errors). SARIMA: ARIMA + seasonal components (P,D,Q,m).

LSTM (Long Short-Term Memory): A type of Recurrent Neural Network designed to capture long-range dependencies in sequential data. Can model complex non-linear patterns but requires large data, is computationally expensive, and is a black box.

💬 "For a SKU with demand only in 8 of the last 36 months, which model would you use and why?"

When to use: MAPE is interpretable (%) but fails with zero/near-zero actual values. MASE is scale-independent and robust to intermittent demand. RMSE penalises large errors more than MAE.

💬 "Why would you prefer MASE over MAPE for comparing forecast accuracy across product categories?"

Star Schema: A central Fact Table (transactions, measurements) connected to multiple Dimension Tables (descriptive attributes) via foreign keys.

Fact Table: Contains quantitative metrics (defect count, units produced, downtime minutes). Dimension Tables: Contain descriptive context (Date, Station, Operator, Product).

Measure vs. Calculated Column: Measures are evaluated at query time in the filter context (dynamic). Calculated Columns are computed row-by-row at data refresh (static, stored in model).

💬 "Why is a star schema preferred over a flat table in Power BI and what is the performance impact?"

-- Defect Rate measure
Defect Rate = DIVIDE(SUM(Defects[defect_count]), SUM(Production[units_produced]), 0)

-- Running total
Running Defects = CALCULATE(SUM(Defects[defect_count]), DATESYTD('Date'[Date]))

-- Rank stations by defect rate
Station Rank = RANKX(ALL(Stations[station_id]), [Defect Rate], , DESC)

-- Prior month comparison
Prior Month Defects = CALCULATE([Total Defects], DATEADD('Date'[Date], -1, MONTH))

-- % Change MoM
MoM Change = DIVIDE([Total Defects] - [Prior Month Defects], [Prior Month Defects], 0)

💬 "What is the difference between CALCULATE and FILTER in DAX, and when would you use CALCULATE?"

KPI Cards (top row): Total Units | Defect Rate % | OEE % | Downtime Hours | FPY %

Trend Charts: Daily defect rate trend (line) | Station-level defect heatmap (matrix) | Downtime by cause (Pareto bar)

Operator View: Skill level vs defect rate scatter | Skill matrix table | Top 10 operators by defect count

Slicers: Date range | Shift | Station | Operator | Defect type

Automation insight: "60 mins/day saved by replacing manual reporting with automated Power BI refresh" — previously team manually compiled station reports; now done automatically via scheduled Power Query data refresh from source database.

💬 "How would you design a dashboard so that a plant manager can identify the top problem station in under 30 seconds?"

Token: The unit of text an LLM processes — roughly a word or sub-word (~4 chars). LLMs have token limits per request.

Embeddings: Numerical vector representations of text that capture semantic meaning. Similar meanings → similar vectors in high-dimensional space.

Context Window: The maximum amount of text (tokens) an LLM can process in one request — determines how much conversation history or document content it can "see".

Temperature: Controls output randomness. Temperature=0: deterministic (most probable next token always chosen). Temperature=1: creative/varied output. Temperature>1: increasingly random/incoherent.

Hallucination: LLMs generating confident but factually incorrect outputs — because they predict statistically likely text, not verified facts.

💬 "What is a hallucination in LLMs and how does RAG help mitigate it?"

RAG (Retrieval-Augmented Generation): Grounds LLM outputs by retrieving relevant documents from an external knowledge base at query time, injecting them into the prompt context. Prevents hallucinations for factual, knowledge-specific tasks. No model training required.

Fine-tuning: Continuing to train an LLM on a domain-specific dataset to adjust the model's weights — embedding domain knowledge into the model permanently. Costly, requires data, better for style/tone/task format than factual grounding.

AI Agent: LLM + Tools + Memory/Context + Planning + Actions. The LLM is the "brain" — it reasons about what tool to call next based on the goal. Tools can be web search, database queries, APIs, code execution.

MCP (Model Context Protocol): A protocol that standardises how AI models connect to external tools and data sources — like USB for AI. Defines Tools (callable functions), Resources (data the model can read), and Prompts (reusable templates).

💬 "When would you choose RAG over fine-tuning for a manufacturing quality assistant?"

Object Detection: Identifies and localises objects in an image with bounding boxes. Object Tracking: Follows detected objects across video frames over time.

YOLO (You Only Look Once): A real-time object detection model that processes the entire image in one pass through the neural network — extremely fast, suitable for production line inspection.

IoU (Intersection over Union): Measures how much a predicted bounding box overlaps with the ground truth. IoU > 0.5 typically = correct detection.

In Manufacturing: False Positive = flagging a good part as defective (cost: rework or scrap of a good unit). False Negative = passing a defective part (cost: downstream defect, warranty claim, safety risk). The trade-off between Precision and Recall is set by the detection threshold.

💬 "In a defect detection CV system, which is worse: high false positives or high false negatives, and how would your answer change for a safety-critical component?"