Lesson 131 — Advanced

Capstone: Deploy on a Managed Platform

Yeh capstone L11 ke saare ideas ko ek kaam mein jodta hai — ek real workload lo aur use managed platform par end-to-end deploy karo: identity se lekar region, quota, gateway routing aur portability kit tak. Interview mein yahi tumhara portfolio-defense story banegi — decision, trade-off aur number, sab apne project se.

⏱ 10 min• Advanced• Portability, Lock-in & Migration

Capstone kya hai — deliverables

Kya hai?

Ek AI workload (jaise ek RAG support-bot ya summarization API) ko kisi ek managed platform par staging se production tak deploy karna — identity, region, quota, gateway, eval gate aur rollback, sab documented.

Kab kaam aata hai?

Jab tumhe dikhana ho ki tum sirf API call nahi kar sakte, balki ek platform decision le sakte ho — residency, cost guardrail aur lock-in ke saath. Yahi interview ka "deploy it properly" round hota hai.

Kahan milta hai?

Har serious AI role ka hiring loop: take-home deploy task, system-design + ops round, ya on-the-job first-week task. Capstone ka demo link resume aur interview dono mein chunkta hai.

Deliverable 1 — Deploy plan: ek page jisme workload, success metric, platform choice, identity, region, quota aur rollout likha ho (neeche template hai).
Deliverable 2 — Live endpoint: staging par chalta endpoint (ya local + exact reproducible commands) with a one-command deploy path.
Deliverable 3 — Numbers: apne hi project ke cost-per-1k-requests, p95 latency aur eval score — borrowed benchmarks nahi.
Deliverable 4 — Portability kit: thin wrapper + golden eval suite + pinned versions + rollback command, taaki provider swap ek config change rahe.

Deploy plan — step order jo actually chalta hai

# capstone deploy plan — teen decision, ek page
workload: "support-bot (RAG over internal docs)"
success_metric: "eval hit@5 apne baseline se better + p95 latency target se kam"
platform: "managed option jo REQUIRED region mein available ho"
identity: "workload identity / role — long-lived key nahi"
region: "data residency se aayega, convenience se nahi"
quotas: "TPM headroom 3x peak + monthly spend guardrail laga di"
gateway: "ek entry point, model id pinned, fallback route defined"
rollout: "eval gate -> shadow -> canary 10% -> full"

Pehle metric, phir platform: "achha lagta hai" decision nahi hota — success metric likho, warna deploy ke baad safai kaise decide karoge.
Region residency se: client data kahan legally padh sakta hai — yeh constraint hai, preference nahi. Uske baad hi performance dekho.
Identity, keys nahi: workload identity ya role-based auth — key leak ka attack surface hi khatam.
Quota + guardrail pehle: Traffic spike ne quota thoka ya spend alarm bajaa — dono planned hone chahiye, incident nahi.
Gateway ek hi: app code directly provider SDK na likhe — gateway/wrapper se route karo, model id pin karo.
Rollout sequence: eval gate (CI mein) -> shadow traffic -> canary -> full. Har step ka rollback ek command se.

Platform rubric — final choice kaise defend karo

Region availability: required region mein kaunsa option available hai — pehla filter, aur aksar final filter.
Quota headroom: current peak ka 3x margin hai ya nahi. Bina headroom ke launch = kal ka outage.
Managed building blocks: platform ka managed vector store / gateway / monitoring use karoge ya third-party? Har choice lock-in ka kharcha le aata hai — knowingly lo.
Cost model: per-token + per-request + egress — teeno jod ke per-1k-requests cost nikalo. Guardrail ke bina budget surprise hai.
Lock-in exposure: kaunse cheezein platform-specific hain (agent graph, index format) — unko list karo, wrapper ke peeche daalo.

Defense rule: Interview mein choice defend karte waqt hamesha do options compare karo with the constraint that decided it (region, quota, cost, managed building block). "Sabse accha tha" ek defence nahi hai.

Portability kit — prod se pehle ready

Thin wrapper: ek request/response format, provider sirf config mein.40 lines se chalu hota hai.
Golden evals: 40-60 cases, category-wise, CI mein chal rahe — yehi migration ka proof hain.
Pinned versions: model id explicit, "latest" kahin nahi — deprecation notice ka bhi owner assign.
Rollback command: previous model id + config ek switch mein — rehearse karo, sirf README mein mat likho.
Migration drill: staging par ek baar provider swap karke dekho — wrapper + evals actually kaam karte hain ya sirf assume kar rahe ho.

portability kit (live endpoint se pehle):
  wrapper        -> ek request/response format, provider config mein
  golden evals   -> 40-60 cases, category-wise, CI mein
  model versions -> pinned ids, latest-word banned
  rollback       -> previous id + config, ek command
  drill          -> staging pe ek baar swap karke verify
sab done = launch ready, warna deploy half-done hai

Order matter karta hai: Portability kit prod ke baad banana sabse mehnga mistake hai — wrapper + evals adoption ke din se banao, tab refactor ka kharcha zero hota hai.

Demo & interview defense — kaise pesh karo

Demo script paanch minute ka rakho: problem -> architecture -> decisions -> numbers -> failure story. Har decision ke saath constraint bolo jo decide karwaya (region, quota, cost). Numbers sirf apne project ke — benchmark ka jhooth interview mein pakda gaya toh trust khatam.

30-second version: "Maine internal docs RAG bot ko managed platform par deploy kiya; region residency se choose kiya, identity-based auth rakha, gateway se pinned model route kiya, aur ek wrapper + eval suite se provider swap ko config-change bana diya."
Trade-off wala sawaal zaroor aayega: "Kya chhoda?" — managed index (lock-in), serverless cold-start (latency), ya cost per request — kuchh toh trade-off admit karo.
Failure story banao: ek cheez jo pehle attempt mein nahi chali (quota hit, eval drop, cold-start latency) aur uska fix — yehi sabse yaad rehne wali baat hai.
Artifacts dikhao: deploy plan page, eval report, cost estimate — screenshots ya repo links, demo ke saath.

Honesty rule: Agar staging hai toh bolo staging hai, simulated traffic hai toh bolo simulated. Overclaim chhota sa hota hai, pakda jaake bada bharosa banta hai.

Exercise

Question: Managed platform choose karte waqt region kis cheez se decide hota hai — performance se ya data residency se?

Question: Long-lived API keys ki jagah modern managed platforms par kaunsa auth pattern recommend hota hai? (role / workload identity)

Question: Provider swap ko config-change banane wali chhoti code layer ka naam? (wrapper / parser)

Question: Rollout sequence ka pehla stage kya hai — canary, shadow ya eval gate?

Common mistakes

Sirf happy path deploy karna: Endpoint chal gaya toh sab done? Quota burst, region outage aur model deprecation bhi plan mein chahiye. Fix: har risk ka drill + rollback rehearse, launch se pehle.
Cost guardrail ke bina ship: Traffic aaya aur bill bhi aaya — bina alarm ke dono surprise hote hain. Fix: monthly spend guardrail + per-1k-requests cost estimate deploy plan ka mandatory part banao.
Evals ko capstone ke baad bhool jaana: Deploy ho gaya toh evals optional? Tab tak koi regression notice hi nahi karega. Fix: eval gate rollout pipeline ke andar — canary se full tak wahi approve karta hai.
Demo mein borrowed numbers: Vendor benchmark paste karke apna result batana — interviewer ek sawaal mein dhool chhaad deta hai. Fix: sirf apne project ke metrics, chhote hon theek hain, asli hon zaroori.

Desi example: Naya ghar shift — managed platform wala plan

Flat shift mein kya karte ho: pehle locality check (region), phir agreement (identity + quotas), phir movers finalize (gateway), saman ki list (workload + evals), aur wapas shift ka plan bhi ready rakhte ho (rollback). Capstone mein bilkul yahi — sirf ghar ki jagah platform hai:

naya ghar (managed platform) ka plan:
  locality   -> region (residency constraint, preference nahi)
  agreement  -> identity + quota + spend guardrail
  movers     -> gateway (ek entry, pinned model, fallback)
  saman list -> workload + golden evals + wrapper
  wapas plan -> rollback command + migration drill
shift bina plan ke = regret; plan ke saath = capstone

Interview angle — yahan se sawaal aate hain

Q: Platform choice kaise defend karoge? — Required region, quota headroom, cost model aur lock-in exposure — constraint ke against do options compare karo.
Q: Deployment mein identity kaise handle ki? — Workload identity / role-based auth, long-lived keys kyun avoid kiye (leak surface).
Q: Rollout stages kya the aur rollback kaise? — Eval gate -> shadow -> canary -> full; rollback = pinned previous id + config, ek command.
Q: Lock-in ka kharcha kitna hai? — Managed index/agent-graph platform-specific; wrapper + golden evals ne swap ko config-change bana diya.
Q: Ek deployment failure batao aur fix? — Apni asli story (quota hit, eval drop, cold-start) — cause, notice karne ka signal, aur permanent fix.

Lesson complete?

Capstone done — tumne managed platform par workload ko identity, region, quota aur portability ke saath deploy kar diya. Yehi story interview mein bolo: decision, trade-off aur number — sab apne project se, koi borrowed nahi.