A free, offline-first interview prep portal: 80 focused GenAI scenarios, a technical field guide, system design, hands-on labs, project defence, and retained broad consulting and technology preparation.
“Analyst” is a level, not one syllabus. Choose the closest track; you can switch anytime.
Move from mechanism → debugging → secure system design → measurable consulting value. Use the cards for rapid review, then practise the prompts, cases, labs, and project-defence drills.
Mutability, generators, typing, exceptions, async I/O, bounded concurrency, timeouts, cancellation, transient retries with jitter, idempotency, strict validation, dependency injection, tests, structured logs, trace IDs, joins, CTEs, windows, null semantics, indexes, and transactions.
Checkpoint: explain why retries need both a policy and an idempotency contract.
Split data by the unit that can leak; choose precision, recall, F1, PR curves, calibration, and thresholds by error cost. Know bias–variance, cross-entropy, perplexity limits, overfitting, distribution shift, and why a simple baseline makes incremental value measurable.
Checkpoint: defend a safety-classifier threshold under alert-capacity constraints.
Explain embeddings; scaled dot-product attention softmax(QKᵀ/√dₖ)V; heads; positional information; residual/normalization blocks; causal masks; encoder, decoder, and encoder–decoder tendencies; tokenizer-dependent cost; decoding; KV cache; and quadratic attention growth.
Checkpoint: draw training vs autoregressive inference without claiming temperature guarantees truth.
Cosine, dot product, Euclidean distance, normalization, compatible dimensions, exact kNN vs ANN/HNSW, metadata filters, BM25, hybrid retrieval, reciprocal-rank fusion, Recall@k, Precision@k, MRR, and nDCG. Re-embed and rebuild when changing embedding spaces.
Checkpoint: explain why exact identifiers often need lexical search.
Acquire → parse/OCR → preserve layout → dedupe → chunk → enrich metadata → embed → index → validate → version. At query time: route/rewrite → retrieve → fuse → rerank → dedupe → pack → answer with citations → abstain. Diagnose retrieval, generation, citation, and freshness separately.
Checkpoint: select chunk size using labelled queries, quality, tokens, p95 latency, and cost—not folklore.
Syntax, schema, semantics, authorization, and business validity are separate layers. The model proposes calls; application code allowlists, authorizes, validates, executes, correlates, limits, audits, and asks for approval. Treat tool results as untrusted data and make side effects idempotent.
Checkpoint: design a refund tool that remains safe after a timeout or injected tool result.
Prefer deterministic workflows when steps are known. Agents need explicit state, permitted tools, budgets, failure transitions, checkpoints, approvals, and terminal conditions. Distinguish working, episodic, semantic, and procedural memory; design consent, provenance, TTL, deletion, tenant isolation, and conflict handling.
Checkpoint: MCP standardizes integration; it does not make a server or tool trustworthy.
Version representative, edge, adversarial, multilingual, and unanswerable cases. Gate ingestion, retrieval, reranking, groundedness, citations, refusal, tool success, safety, latency, and cost separately. Calibrate LLM judges against people; compare paired baselines; trace every component with privacy controls.
Checkpoint: define a release gate that catches a retrieval regression hidden by fluent answers.
Direct and indirect prompt injection, sensitive disclosure, supply-chain and poisoning risks, unsafe output handling, excessive agency, vector/embedding weaknesses, misinformation, and unbounded consumption require defence in depth. Enforce ACLs before context assembly; minimize and redact traces.
Checkpoint: delimiters and secret prompts are hints, not security boundaries.
Use RAG for fresh/private knowledge, citations, updates, and deletion. Fine-tune behaviour, style, classification, formatting, or repeated tool patterns only after a prompt/retrieval baseline and held-out eval. PEFT can reduce training/storage cost; it does not turn weights into an auditable database.
Checkpoint: name the failure layer before choosing a remedy.
Clarify current Microsoft Foundry vs classic resource/API generations. A defensible option combines Entra ID, managed identity/RBAC, Azure OpenAI/Foundry deployments, Azure AI Search hybrid retrieval, governed sources, justified compute, Key Vault, private networking, APIM, Application Insights, CI/CD eval gates, and rollback.
Checkpoint: verify names, routes, API/SDK versions, preview status, regions, quotas, and pricing that morning.
Start with users, answerability, sensitivity, traffic, SLOs, action risk, quality, and budget. Use stateless APIs, external state, queues, backpressure, scoped caches, model routing, circuit breakers, rate limits, load tests, versioned artefacts, canaries, and observable rollback.
Checkpoint: allocate the latency budget across gateway, retrieval, reranking, model, and tools.
Answer with claim → mechanism → measurement → trade-off → production control. For behavior, use Situation → your scope → actions and alternatives → measured result → learning. State baselines, failures, evidence, ownership boundaries, client impact, rollout, and responsible-AI controls.
Checkpoint: replace “we used RAG” with an evaluated decision and a quantified trade-off.
Use each as a 3–7 minute drill. The criteria guide coverage; they are not scripts or claims about an employer’s exact questions.
Cover tokens/embeddings, Q/K/V attention and scaling, heads, positions, feed-forward/residual/normalization, causal masking, training vs inference, and long-context cost.
Compare layout/heading/table-aware parsing, parent–child chunks, overlap, metadata, OCR validation, and an empirical sweep.
Separate retrieval, packing, conflicting/stale context, instruction, model, decoding, and citation failure; propose layer-specific evals.
Use identifier, jargon, and paraphrase examples; discuss filters, RRF, first-stage recall, reranker latency, and measured trade-offs.
Discuss fresh/private knowledge, behaviour adaptation, citations/deletion, data quality, PEFT, held-out evals, cost, and combined use.
Define schemas, authorization, deterministic limits, idempotency, approval, audit, injection resistance, budgets, error states, and task-success evals.
Explain host/client/server, JSON-RPC, resources/prompts/tools, capability negotiation and consent; interoperability does not guarantee trust.
Cover memory types, consent, provenance, scope, retrieval, conflict, TTL, deletion/export, tenant isolation, and summary loss.
Dataset, retrieval/answer/citation/refusal/safety metrics, human and calibrated LLM judges, paired baseline, CI gate, monitoring, and slices.
Instrument first; prune context, right-size models, parallelize I/O, stream, cache safely, batch offline work, bound agents, and protect quality.
Clarify Foundry generation, identity/RBAC, Search, networking, governed sources, compute, Key Vault, APIM, monitoring, CI/CD, and data policy.
Give direct/indirect examples, business impact, why secrecy/delimiters fail, and capability-based defence in depth.
Show reproduction, traces, hypothesis, fix, tests, rollback, and prevention for async blocking, retry storms, mutable state, or fan-out.
Use evidence-rich STAR: own scope, quantify failure, find root cause, communicate, change the system, and add a regression control.
Discovery, baseline, source readiness, use-case priority, risk, prototype, acceptance metrics, pilot, governance, ROI, change management, and exit criteria.
For every case: clarify requirements, then draw both the online request flow and the offline data/evaluation flow.
Brief: 50,000 employees; English/Hindi; SharePoint/PDF; HR ACLs; p95 < 4 s; citations. Cover: identity propagation, security trimming, layout/OCR, multilingual hybrid retrieval, freshness, reranking/citations, abstention, private Azure option, eval slices, SLO/cost, ACL-scoped caches.
Brief: compare deal-room files across entities/years. Cover: engagement isolation, file validation, table/layout extraction, entity/date metadata, decomposition, parent–child retrieval, page evidence, no cross-deal cache, retention/legal hold, human review, quantitative accuracy, bounded spend.
Brief: answer products, fetch status, block cards—never transfer funds. Cover: deterministic router, RAG, scoped tools, reauthentication, confirmation, absence of transfer capability, idempotency, budgets, injected tool output, support paths, audit, policy evals.
Brief: explain code, find ownership, suggest patches, open a PR after approval. Cover: symbol-aware plus hybrid retrieval, commit metadata and ACLs, sandbox tests, supply-chain scan, diff review, approval, secret safety, untrusted integrations, acceptance metrics.
Brief: English/Hindi/Marathi, low-bandwidth mobile, frequent policy change, accessible. Cover: authoritative sources, multilingual/hybrid search, concise streaming, citations/escalation, freshness SLO, safety, offline constraints, and evaluation by language/literacy/geography.
Brief: global, 99.9%, p95 < 2.5 s, strict unit economics. Cover: capacity, autoscaling, admission control, bounded retrieval/reranking, model escalation, scoped/versioned caches, streaming, circuit breakers, regional strategy, graceful degradation, load and cost tests.
Brief: 40 clause types, strict schema, cited spans, lawyer review. Cover: queue, OCR/layout, constrained extraction, deterministic validation, calibrated evidence, exact spans, clause evals, review queue, lineage/reprocessing, residency/deletion, batch economics.
Brief: shared guardrails, tools, and monitoring for many business units. Cover: control/data planes, Foundry-generation clarity, project boundaries, identity/RBAC, approved catalogue, tool gateway, policy as code, network isolation, eval/red-team gates, tracing, quotas, registry, kill switch, rollback.
Each lab should produce a small repo/notebook, README, tests/eval data, measured results, and a two-minute explanation—without secrets or proprietary data.
Requirements/risks → Quality/SLO/cost → Architecture online+offline → Data/index/freshness/ACLs → Evals/gates → Security/privacy → Capacity/cache/cost/failures → Observability/rollout/rollback.
Reproduce and define expected behaviour → inspect input-to-output lineage → isolate the first bad boundary → compare version/config/data → fix the owning layer → add a regression case and production signal.
Situation (brief) → Task and your scope → Actions and alternatives → Result with metrics → Learning. Include stakeholder communication and what you would change.
Architecture story: async Python orchestration coordinates task interpretation, notebook-cell generation, Synapse execution, Spark/Livy polling, log classification, selective cell patching, validation evidence, and Azure DevOps branch/PR automation—with human approval before merge.
Architecture story: FastAPI gateway, governed retrieval with metadata/ACL filters and citations, role-aware tool calling, Redis-backed task/session state, explicit approval for production-changing actions, and correlated traces.
For GenAI / Agentic AI roles: prepare across Python and SQL, ML/DL and transformers, retrieval/RAG, tools and bounded agents, evaluation/security, Azure architecture, coding, system design, project ownership, and consulting communication. A representative—not guaranteed—loop is screen → fundamentals → hands-on → architecture → project deep-dive → behavioral/client round.
There is no single universal “Deloitte paper.” Deloitte India and Deloitte US-India Offices (USI) are distinct organizations, and the sequence changes by practice, role, campus, and hiring campaign. Use this as a preparation map—not a promised process.
Official Job-specific profile or campus placement route; recruiters and hiring managers review fit.
Observed Some fresher technology drives report aptitude, verbal/reasoning, CS MCQs, and coding. Other practices report business or domain assessments instead.
Projects and resume depth plus role fundamentals—coding/SQL/OOP for software, Excel/Python/statistics for data, accounting/controls for audit or risk, and platform depth for experienced hires.
Official guidance Case/scenario interviews assess clear thinking, assumptions, analysis, recommendations, and impact—not one memorized “right answer.”
Resume ownership, communication, leadership, teamwork, ambiguity, priorities, client judgment, and genuine STAR examples. This may be combined with another interview.
Some campaigns add a manager, Director/Partner, or HR round; others conclude after a combined interview. Offer and onboarding follow only after the actual process is complete.
Frequently reported: timed aptitude/verbal/reasoning + CS/technical MCQs + two coding problems, then a technical or combined interview; a separate manager/HR round appears in some—not all—drives.
Prioritize: arrays/strings, SQL, OOP, DBMS/OS/network basics, projects, and behavioral scenarios.
Some campus tracks report an assessment, business discussion or GD, then senior interviews focused on projects, data/GenAI, commercial thinking, and scenarios rather than a coding-heavy funnel.
Prioritize: structured communication, business application, project ownership, cases, and senior-panel maturity.
Expect deeper probing of end-to-end delivery, design choices, troubleshooting, stakeholder workshops, testing, risk/quality, and measurable impact. Generic puzzle preparation has lower return.
Prioritize: turn every important JD bullet into one defensible project story.
Python/SQL, ML and transformers, embeddings and hybrid retrieval, RAG, schemas and tool calling, bounded agents/MCP/memory, evals/observability, security, Azure, cost/latency, and defensible production projects.
Coding aloud, arrays/strings, OOP, DBMS, OS, networks, SQL, project architecture, tests, and edge cases.
SQL, Excel, Python/pandas, statistics, model evaluation, cleaning, dashboards, and business interpretation.
Clarify, structure, analyse, synthesise, recommend, and state risks and next steps; add commercial awareness.
Risk-control mapping, security basics, named platforms, implementation lifecycle, troubleshooting, and documentation.
Role-relevant accounting and audit fundamentals, evidence, controls, ethics, scenarios, and communication.
Structure behavioral answers using Situation-Task-Action-Result.
Revenue (Price × Volume) vs Costs (Fixed + Variable). Segment by product/region/channel.
Market size → Growth → Competition → Customer segments → Capabilities → Entry mode → Financials.
Top-down (TAM→SAM→SOM) or Bottom-up (units × price). State assumptions explicitly.
Cost-plus, value-based, competitive benchmarking. Consider elasticity and willingness-to-pay.
Strategic rationale → Standalone value → Synergies → Risks → Valuation → Integration.
Bottleneck analysis, capacity planning, lean/six-sigma principles, automation ROI.
Write solutions below. Code runs locally in a disposable worker with a two-second limit; output appears in the console area. Do not paste untrusted code.
Draft a query, then compare it with a reference solution. SQL dialects vary; no query is sent anywhere.
| Topic | Formula |
|---|---|
| Speed-Distance-Time | S = D/T; Relative speed (same dir) = |s1-s2| |
| Work | Combined rate = 1/A + 1/B; Time = 1/(combined rate) |
| Percentage | %change = (diff/original)×100 |
| Profit/Loss | Profit% = (SP-CP)/CP × 100 |
| SI/CI | SI=PRT/100; CI=P(1+R/100)^T - P |
| Probability | P(A∪B)=P(A)+P(B)-P(A∩B) |
| Permutation | nPr = n!/(n-r)! |
| Combination | nCr = n!/[r!(n-r)!] |
| Clause | Use |
|---|---|
| WHERE | Filter rows before grouping |
| HAVING | Filter groups after GROUP BY |
| PARTITION BY | Window function grouping |
| COALESCE | First non-null value |
Start with the dedicated GenAI plan; every GenAI day ends with spoken recall and an error log. Broad assessment plans remain available.
Important: Aptivra is an independent, free educational resource. It is NOT affiliated with Deloitte, any Big-4 firm, or any recruitment agency. No proprietary or confidential material is used.
Recruitment processes change frequently and can differ by business unit, campus, geography, role, and candidate. Aptivra does not claim that any roadmap or test format is current or universal. Verify the role description and official candidate communication on the day you apply.
Official sources define guidance and organizational context. Observed sources are individual candidate reports—useful for patterns, never guarantees.
Stable concept principles still require domain evaluation. Product/version-sensitive links must be rechecked for current API versions, names, limits, preview/GA status, regions, and pricing.
All 208 questions are representative original compositions, including 80 GenAI scenarios numbered 129–208. No proprietary, recalled, or leaked live-test material is reproduced. Aptivra makes no claim about exact interview questions, rounds, timing, or selection patterns. Difficulty and coverage are approximations; the current job description, invitation, and official documentation remain authoritative.