Lesson 168 — Advanced

Interview strategy & role positioning

Interview ka score sirf knowledge se nahi banta — kaunsa role, kaunsa sawaal, kis frame se — isi decide karte hain. Is lesson mein tum seekhoge ki GenAI interviews ab theory ratne se kis direction mein shift ho gaye hain, JD ko kaise annotate karte hain ki roadmap ke kaunse areas asli mein test ho rahe hain, AI Engineer / GenAI DS / traditional DS roles ka emphasis-farak, API-hosted vs self-hosted models ko kaise defend karte ho, aur ek personal prep plan jo tumhari asli weaknesses pe attack karta hai.

⏱ 10 min• Advanced• Basic computer usage

GenAI interviews kaise badle — theory se production tak

Kya hai?

Interview strategy ka matlab hai apni limited taiyari ko sahi sawaal pe lagana. GenAI interviews ab definition-wale sawaal se hat kar system design, debugging, architecture aur production scenarios pe chal gaye hain — "50,000 concurrent users pe RAG kaise chalayein" poochha jaata hai, "embedding kya hai" nahi.

Kab kaam aata hai?

Jab bhi tum koi naya role apply karo ya internal transfer lo. Pehle din se hi — JD padhte waqt, prep plan banate waqt, aur interview ke andar jab interviewer jaan-boojh kar ambiguity daale. Frame pehle se tay ho toh sawaal dar nahi lagta.

Kahan milta hai?

Har GenAI hiring loop mein: HR screen (role fit), technical screen (ek project deep-dive), system design round (architecture defend), aur debugging round (incident scenario). Round ka naam alag ho, sawaal ke pattern yehi hain.

Shift ka mechanism ye hai: model ab API mein bikta hai, toh koi bhi banda "LLM use kiya" bol sakta hai. Isliye companies ka filtering power un sawaalon mein hai jinhe cheat nahi kar sakte — tumhara system down ho jaaye toh kya karoge, quality giri toh kaise dhoondhoge, 10x traffic aaye toh kya tootega. Ye sab production ka experience maangte hain, definition ka nahi.

Rule of thumb: Interview mein "mujhe pata hai" prove nahi karna — "main soch sakta hoon" prove karna hai. Har jawab mein dikhao: pehle assumptions, phir structure, phir trade-off, aur end mein ek concrete example. Ye chaar step hi strategy hai.

JD ko annotate karo — kya test ho raha hai

JD line (extract)                          -> roadmap area jo test ho raha hai
"design and operate RAG pipelines"         -> L4 RAG + L10 LLMOps (ingestion, retrieval, evals)
"agents, tool use, MCP integrations"       -> L5/L6 agents + L11 platforms
"deploy models on cloud / GPU"             -> L11 managed platforms / L12 serving
"monitor quality, cost, latency"           -> L10 observability + SLOs
"evaluate and fine-tune models"            -> L13 fine-tuning + L14 evals
"Python, CI/CD, containers"                -> engineering layer — coding bhi poochha jaayega

rule: 3-4 core areas -> unhi pe depth; baaki sab ke liye surface awareness

JD ke verbs dekho (design, operate, deploy, evaluate, monitor) — wahi batate hain round ka shape. Operate zyada likha hai toh on-call style debugging round pakka hai.
Tools ki list ko goal samjho, checklist nahi. "LangChain/LlamaIndex" ka matlab hai "RAG orchestration aani chahiye" — library ki API ratni nahi.
Jo area JD mein baar baar aaye, wahi deep-dive round hoga — uska ek chhota design taiyaar rakho, 5 minute mein whiteboard pe.
Missing cheez bhi signal hai: evals ka zikr hi nahi toh interview mein pooch lo "yahan quality kaise naapi jaati hai?" — ye sawaal dikhata hai ki tum system ke baare mein sochte ho.

JD ka trick: Listed har tool seekhne ki koshish mat karo — time khatam hoga aur depth kahin nahi. 3-4 core areas chuno, unmein itna do ki interviewer ko lage "isne actually chalaya hai", baaki sab ke liye seedha bol do "ye maine nahi chhua, par ye seekhunga".

AI Engineer vs GenAI Data Scientist vs Data Scientist — emphasis ka farq

                        AI ENGINEER           GENAI DATA SCIENTIST    DATA SCIENTIST
core emphasis            production systems    model/prompt behaviour  statistical modelling
roz ka kaam             services, pipelines  experiments, evals      features, models, metrics
interview mein zyada    system design, ops,  eval design, error      classical ML, statistics,
                        debugging, scaling   analysis, adaptation    data quality, leakage
deliverable             shippable service    defensible decision     trained model + report

Ye teeno alag disciplines nahi hain — same concepts hain, alag-alag emphasis. Fark center of gravity ka hai.
AI Engineer ka interview engineering layer pe zyada hota hai: APIs, inference, LLMOps, observability, networking, cloud, reliability, deployment.
GenAI DS ka zyada: prompt vs RAG vs fine-tune decisions, eval design, failure analysis, error cases ka taxonomy.
Traditional DS ka ab bhi poochha jaata hai: statistics, leakage, metrics, classical models — L16 wala layer interview se gaya nahi hai.
Apna answer role ke hisaab se weight do: same project — engineering role pe "kaise scale kiya", DS role pe "kaise measure kiya" bolo.

Positioning: Job title pe mat jaao, roz ke kaam pe jaao. "Main AI Engineer hoon" kehne se pehle bolo "mera roz ka kaam services banana, deploy karna aur incidents sambhaalna hai" — interviewer khud label laga dega, aur sawaal bhi usi taraf jaayenge jahan tum strong ho.

API-hosted vs self-hosted — 6 axis pe defend karo

axis            API-hosted                    self-hosted open-weight
capability      top models, update milte     model chunna tumhara kaam
cost            per-token, variable           GPU fixed cost, high usage pe sasta
latency         network + provider dependent  apne control, batching apna
control         params/features provider pe   poora control (quantization, kernels)
privacy         data provider ke paas jaata   data bahar nahi jaata
ops burden      lagbhag zero                  serving, scaling, patching sab tumhare

Decision rule: privacy ya control non-negotiable ho toh self-hosted; warna API se shuru — complexity tab jab demand dikhaaye.
Hybrid pattern: public workload API pe, sensitive workload self-hosted — do pipelines, ek hi evaluation harness, taaki comparison honest rahe.
Cost ka defence: high steady-state traffic pe per-token bill GPU fixed cost se cross hota hai. Break-even point nikalo, "mehnga-lagta-hai" se nahi.
Interview mein answer tab strong hota hai jab dono taraf ka dard bolo: "API ka vendor lock-in aur self-hosted ka on-call burden — dono real hain, farq workload ke numbers mein hai."

Honest comparison: Dono taraf ke numbers maango: traffic kitna, tokens per request kitne, latency budget kya, data kahan ja sakta hai. Inke bina koi bhi side ka answer opinion hai, engineering nahi. Interviewer ko farq dikhta hai.

Personal prep plan — weakness se shuru

week_plan:            # frame hai — apne numbers apne hisaab se daalo
  day_1: mock round attempt (timed, aloud) -> gaps note karo evidence ke saath
  day_2_3: sabse bada gap -> us ek area ki dobara padhai + ek chhota build
  day_4: project defense story likho (problem -> decision -> number)
  day_5: dobara mock -> same sawaal, ab structure ke saath
  rotating_focus: system_design | debugging | project_deep_dive | behavioral
scorecard_per_attempt:
  assumptions_stated: true/false
  structure_followed: true/false
  tradeoffs_named: true/false
  numbers_real: true/false
  honest_iknow_not: true/false

Mock attempt hi asli diagnostic hai — padhne se weakness dikhti nahi, bolne (aur record karke sunne) se dikhti hai.
Har mock ko rubric pe score karo: assumptions, structure, trade-offs, real numbers, honesty. Improvement isi scorecard se dikhegi, feeling se nahi.
80/20 rule: sabse kamzor 2 areas pe 80% time do, apne favourite topics pe nahi — interview wahi poochega jo tum chup kar rahe ho.
Har din ek project story dobara bolo — jaise interviewer saamne baitha ho, aloud, bina notes ke. Ratna nahi, bolna practice hai.

Imaandari rule: Jo nahi aata wo seedha bol do aur saath mein batao ki kaise seekhoge: "Ye maine production mein nahi chhua, par iska mechanism ye hai aur main kaise start karunga." Interviewer bluff se zyada self-awareness ko score karta hai.

Exercise

Question: GenAI interviews ka centre ab kis cheez pe hai — definition ratna ya production scenarios?

Question: JD mein listed har tool seekhne ke bajaye uske peeche kya chhupa hota hai? (ek shabd)

Question: Privacy non-negotiable hai — API-hosted ke bajaye kaunsa option defend karoge?

Question: Apni weakness dhoondhne ka sabse sachcha tareeka kya hai — articles padhna ya timed mock attempt?

Question: AI Engineer aur traditional Data Scientist ka farq kis shabd mein — AI Engineer ka centre kis layer pe hai?

Common mistakes

Har listed tool ka demo banana: JD mein 15 tools hain aur tum sab seekhne nikal pade — time khatam, depth kahin nahi. Fix: 3-4 core areas chuno, unmein itna do ki mechanism clear ho.
Theory rat kar interview jaana: Definitions smooth, par "ab design karo" pe freeze. Fix: roz ek scenario mock karo — aloud, timer ke saath, aur khud ko rubric pe score karo.
Sab roles ko ek samajhna: DS wale round mein ops pooch liya ya ulta — positioning hi nahi bani, toh sawaal bhi galat jagah ja rahe hain. Fix: role ke roz ke kaam padho, phir apni story unhi pe align karo.
Favourite projects hi bolna: Interviewer us project pe sawaal karega jo JD se match karta hai, jo tumhara sabse purana ya sabse impressive nahi hai. Fix: JD se 2-3 projects chuno, unhi ki poori taiyari karo.

Desi example: Kirane ki dukaan ka stock — JD hi demand hai

Prep ka matlab inventory bharna nahi — demand dekh kar stock laana hai. Jo cheez customer (interviewer) maang raha hai wahi rakho, warna godown bhar gaya aur sale zero. Isliye pehle JD padho, phir planning karo:

demand (JD)    -> "RAG + LLMOps + cloud deploy"
inventory      -> L4 + L10 + L11 ki depth (3-4 core areas)
dead stock     -> wo tools jo JD mein kahin bhi nahi
restock plan   -> mock attempt -> gap note -> targeted lesson
sale feedback  -> har mock ka scorecard (structure, trade-offs, numbers)

Interview angle — yahan se sawaal aate hain

Q: GenAI interviews theory se kis direction mein shift hue? — System design, debugging, architecture aur production scenarios ki taraf; "kya hai" se "kaise chalayenge" tak.
Q: JD ko kaise padhoge? — 3-4 core verbs/areas nikaalo aur unhi pe depth banao; listed tools ko goals samjho, checklist nahi.
Q: AI Engineer vs GenAI Data Scientist ka farq? — Emphasis: engineering layer (services, ops, scaling) vs model behaviour (evals, adaptation, error analysis). Alag disciplines nahi hain.
Q: API-hosted vs self-hosted kab chunoge? — Privacy/control non-negotiable toh self-hosted; warna API se shuru. Break-even traffic pe cost compare karo, guess nahi.
Q: Weakness kaise identify karoge? — Timed mock attempt karke rubric pe score karo; padhne se nahi, bolne se gaps dikhte hain.

Lesson complete?

Strategy ka base ready — ab har sawaal ka frame tumhare paas hai. Agla lesson — Production RAG Interview Scenarios — mein is frame ko ek real whiteboard design pe lagayenge.