AI ইন্টারভিউ প্রস্তুতি

Stage 1: Coding (DSA + ML)

• LeetCode Top 150 — easy/medium-এ fluent হন।
• NumPy/Pandas — vectorized solution।
• "Implement X from scratch" — k-means, logistic regression, attention।
• PyTorch live coding — DataLoader, training loop।

# Classic: scaled dot-product attention from scratch
import torch, math
def attention(Q, K, V, mask=None):
    d_k = Q.size(-1)
    scores = Q @ K.transpose(-2,-1) / math.sqrt(d_k)
    if mask is not None: scores = scores.masked_fill(mask==0, -1e9)
    return torch.softmax(scores, dim=-1) @ V

Stage 2: ML Breadth

Quick-fire concept question — ১-৩ মিনিটে clear উত্তর।

• Bias-variance tradeoff।
• Overfit vs underfit — symptom ও remedy।
• Regularization — L1/L2, dropout, early stop।
• Precision/Recall/F1, ROC-AUC, PR-AUC কখন কোনটা।
• Cross-validation strategy (stratified, group, time-series)।
• Gradient descent variants।
• Batch norm vs layer norm vs RMS norm।
• Encoder vs decoder vs encoder-decoder transformer।
• LoRA, QLoRA, PEFT।
• RAG vs fine-tuning কখন কী।

Stage 3: ML Depth

একটা topic-এ deep dive — paper details, math, trade-off।

• "Transformer-এ positional encoding কেন লাগে? RoPE vs ALiBi?"
• "BERT pretraining objective derive করুন।"
• "Diffusion forward/reverse process explain করুন।"
• "DPO vs PPO RLHF-এ?"

Stage 4: ML System Design (সবচেয়ে গুরুত্বপূর্ণ)

Senior role-এ এটাই deciding factor। Framework:

1. Clarify — user, scale, latency, budget।
2. Metric — online (CTR, retention) + offline (AUC, NDCG)।
3. Data — source, labeling, leakage, freshness।
4. Feature — feature store, online/offline parity।
5. Model — baseline → complex, candidate gen + ranker।
6. Training — schedule, retraining frequency।
7. Serving — batch/online, cache, fallback।
8. Monitor — drift, performance, A/B test।

Stage 5: Behavioral (STAR)

প্রতিটি story-তে: Situation → Task → Action → Result। নিজের ৬-৮টি story আগেই তৈরি রাখুন।

• "একটা challenging project বলুন।"
• "Conflict কীভাবে handle করেছেন?"
• "Failed project থেকে কী শিখেছেন?"
• "কেন এই company?"

Resource Stack

• Book: Designing ML Systems (Chip Huyen), ML Interviews (Khang Pham)।
• Course: Educative "Grokking the ML Interview", Stanford CS329S।
• Mock: interviewing.io, Pramp, friend pair।
• Blog: Eugene Yan, Chip Huyen, Vicki Boykis।

সারসংক্ষেপ