Interview Preparation — মেশিন লার্নিং

Hook — Skill ≠ Offer

তুমি ভালো model বানাতে পারো, কিন্তু interview process আলাদা game — coding, ML breadth, ML depth, system design, behavioral — পাঁচটা stage prepare করতে হবে।

Typical ML Interview Loop

stages

1. Recruiter screen      — 30 min, fit + comp
2. Coding (DS/Algo)      — 1-2 round, LeetCode med
3. ML Breadth            — quick fire concepts
4. ML Depth              — past project deep dive
5. ML System Design      — design a recommender etc.
6. Behavioral / Bar      — STAR stories, leadership
7. Team match / Offer

Coding Round

LeetCode top-150 + Blind 75 medium।
Pattern: array/hashmap, two-pointer, sliding window, BFS/DFS, DP basic, heap।
Talk through brute → optimized → complexity।
Python + clean naming + edge case।
Mock interview: Pramp, interviewing.io।

ML Breadth Topics

Bias-variance, overfitting, regularization।
Cross-validation, train/val/test split mistake।
Logistic vs Linear, SVM, tree-based।
PCA, t-SNE, clustering।
Gradient descent variants, learning rate schedule।
Metric: precision/recall/F1, ROC vs PR, calibration।
Imbalanced classes — SMOTE, class weight, threshold।
Encoder/Decoder, attention, transformer basics।

Project Deep Dive — STAR + Why

answer-frame

Situation  — context, scale, constraints
Task       — your specific responsibility
Action     — what you did, alternatives considered
Result     — quantified impact (X% lift, $Y saved)
Why        — why this choice over alternatives
Now what   — what would you do differently today

Interviewer ‘why not other approach?’ জিজ্ঞেস করবে — ৩টা alternative ready রাখো।

ML System Design Framework

6-step

1. Clarify problem & scope (latency, scale, metric)
2. Define success metric (offline + online)
3. Data — sources, labels, sampling, leakage
4. Model — baseline → upgrade path
5. System — training pipeline + serving + monitoring
6. Tradeoffs, failure modes, future iteration

Common asks: news feed, ads CTR, recommender, search, fraud, chatbot, RAG, ride ETA।
Draw diagram — user → service → model → store।
Mention monitoring + drift + retraining cadence।

Behavioral — Story Bank

Conflict with teammate / PM।
Failed project + what you learned।
Ambiguous problem — কীভাবে scope করলে।
Influence without authority।
Mentoring / receiving feedback।
৬–৮ story tailor — STAR এ practice।

Offer & Negotiation

কখনো প্রথম number বলো না — ‘what's the range?’।
Total comp = base + bonus + equity + sign-on।
Competing offer (real one) — biggest lever।
Levels.fyi, glassdoor দিয়ে benchmark।
Always negotiate — companies expect করে।
Written offer এর আগে কিছু sign করবে না।

8-Week Prep Plan

weekly

Wk 1-2: LeetCode patterns + ML breadth flashcards
Wk 3-4: 2 deep-dive projects polished + STAR draft
Wk 5-6: ML system design (1 case study/week)
Wk 7  : Mock interviews (3-5)
Wk 8  : Behavioral polish + comp research + apply

Summary

এক নজরে

ML Interview = Coding + Breadth + Depth + System + Behavioral। STAR + tradeoff বলতে পারলে অর্ধেক জয়। Mock + negotiation skip করো না।