← WhatTNext AI सर्टिफ़िकेशन परीक्षा गाइड

सर्टिफ़ाइड AI प्रोफ़ेशनल

कोर्स और परीक्षा अभी अंग्रेज़ी में हैं। उनके आसपास के पेज हिंदी में हैं।

यह किसे प्रमाणित करता है

A Certified AI Professional evaluates and leads the use of AI in an organisation. They choose between prompting, retrieval, fine-tuning and agent designs; plan evaluation and red-teaming; identify security, privacy and fairness risks and the controls that address them; weigh cost, latency and deployment trade-offs; and build the business case, policy and change plan for adoption, in line with India's DPDP Act 2023, the EU AI Act, the NIST AI Risk Management Framework and India's AI Governance Guidelines. The exam is built around published case studies. It does not certify the ability to train models or write production code.

किसे देनी चाहिए

सुझाया गया अनुभव

Associate-level skills (the Certified AI Associate course or equivalent) and at least one year contributing to or leading AI-enabled work in product, operations, analytics, risk, compliance, IT or consulting. Programming is not required, but you should be able to read a simple system diagram and a cost table.

परीक्षा

  • 120 मिनट में 50 अंक वाले सवाल
  • 25% ज्ञान, 75% व्यावहारिक क्षमता
  • एक व्यावहारिक कार्य, स्कोर का 20%, नतीजा परीक्षा ख़त्म होते ही
  • 5 अतिरिक्त परीक्षण सवाल जो गिने नहीं जाते, ताकि नए सवाल परखे जा सकें
  • एक जवाब, कई जवाब (सब सही हों तभी अंक) और क्रम वाले सवाल, कई असली प्रॉम्प्ट और AI जवाबों पर बने
  • नतीजा: एक तय पास मानक पर पास या फ़ेल, हर सेक्शन के लिए "मानक पूरा" या "सुधार चाहिए"
  • ऑनलाइन आँकी गई, एक बार में एक सवाल, शुरुआत में वैकल्पिक वेबकैम पहचान जाँच
कोर्स लीजिए $20 / ₹799

सेक्शन और भार

सेक्शनभारजाँचे जाने वाले कौशल
Model foundations and architecture choices14%
  • Explain pre-training, instruction tuning and preference alignment, and how each shapes model behaviour.
  • Choose between prompting, retrieval-augmented generation (RAG), fine-tuning and a smaller specialised model for a stated requirement.
  • Explain embeddings and vector search, and evaluate chunking, metadata and versioning choices in a RAG design.
  • Compare open-weight and closed (API) models on control, data location, capability and operating effort.
  • Analyse how retrieval quality, context length and model choice each affect answer quality.
Agents, tool use and system design16%
  • Describe how an agent plans, calls tools and uses their results, and where each step can fail.
  • Evaluate an agent design for least privilege, human approval points and safe stopping conditions.
  • Identify agent failure modes: loops, wrong tool, compounding errors, over-reach and silent failure.
  • Design guardrails: input and output checks, tool allow-lists, step and spend limits, and audit logs.
  • Decide when a fixed, deterministic workflow is better than an autonomous agent.
  • Assess risks across the AI value chain: third-party models, plug-ins, data sources and vendors.
Evaluation, red-teaming and quality assurance20%
  • Build an evaluation set that reflects real use, edge cases, languages and known failure modes.
  • Select metrics suited to the task: correctness, groundedness, retrieval recall, refusal behaviour, latency and human ratings.
  • Design human review and sampling proportionate to risk, and guard against automation bias.
  • Plan red-team tests for prompt injection, jailbreaks, data leakage and harmful output.
  • Interpret evaluation results, including differences between user groups, and make a go/no-go decision.
  • Set up monitoring and regression testing after launch, including when a model or prompt changes.
Security, risk and responsible AI governance20%
  • Identify LLM-specific security risks — direct and indirect prompt injection, sensitive-information disclosure, excessive agency — and matching controls.
  • Apply the NIST AI RMF functions (Govern, Map, Measure, Manage) and the Generative AI Profile's risk list to a use case.
  • Assess a use case against DPDP Act 2023 duties and the DPDP Rules 2025 timeline: notice, consent, purpose limitation, security safeguards, breach notification, children's data and retention.
  • Determine whether an EU AI Act prohibition, high-risk category or transparency duty applies to a described system.
  • Apply the India AI Governance Guidelines' seven sutras and any sector guidance (e.g. the RBI's FREE-AI report) to a deployment decision.
  • Evaluate fairness and bias risks, including proxy variables, and choose mitigations.
  • Define accountability: owners, documentation, incident response and grievance redressal.
Cost, latency, deployment and model selection10%
  • Estimate running cost from tokens per request, request volume and price per million tokens.
  • Evaluate trade-offs between model size, quality, latency and cost using evaluation results.
  • Choose techniques that cut cost or latency: routing to smaller models, caching, batching and shorter context.
  • Compare deployment options — vendor API, cloud-hosted, self-hosted open-weight — for data location, control and operating load.
AI strategy, adoption and change leadership20%
  • Prioritise AI use cases by value, feasibility, risk and data readiness.
  • Build a business case with a baseline, measurable outcomes, total cost of ownership and risk-adjusted benefits.
  • Design a pilot with success criteria, a comparison baseline and exit criteria agreed in advance.
  • Plan change management: affected roles, consultation, training and communication.
  • Write or evaluate an AI acceptable-use policy and a role-based AI-literacy plan.
  • Define post-launch impact measures and the conditions to scale, pause or stop.

यह स्तर किन ढाँचों के अनुरूप है

हमने पाठ्यक्रम बनाने में इन सार्वजनिक ढाँचों का इस्तेमाल किया। ये इस सर्टिफ़िकेशन का समर्थन नहीं करते।

नमूना सवाल

ये शैली और स्तर के उदाहरण हैं। ये परीक्षा में नहीं आते।

1.

Tejomaya Bank's product notes and circulars change every week, and relationship managers must be able to see which document an answer came from. Which approach best fits the policy-assistant requirement?

  1. Fine-tune a large model every month on all 2,600 documents, so that it memorises the bank's policies and answers from memory.
  2. RAG over a curated, versioned document index that returns cited passages.
  3. Place all 2,600 documents in the prompt of every request, using the model with the largest context window available.
  4. Train a new language model from scratch on the bank's internal documents, so that no outside model is involved.

जवाब: B

Frequently changing content that must be cited fits retrieval: documents can be updated or withdrawn in the index the same day, and each answer can point to its source. Fine-tuning bakes in stale content and can't cite; putting everything in the prompt is costly and slow and degrades focus; training from scratch is out of proportion.

2.

In the OWASP Top 10 for LLM Applications (2025), which risk describes an LLM-based system taking damaging actions because it has been given more functionality, permissions or autonomy than its task needs?

  1. Prompt injection
  2. Excessive agency
  3. Unbounded consumption of compute and money
  4. Supply-chain compromise of third-party components

जवाब: B

Excessive Agency (LLM06:2025) covers excessive functionality, permissions and autonomy. Prompt injection is often the trigger, but limiting agency is what bounds the damage.

3.

The steering committee fears that a customer's dispute attachment could contain hidden text telling Tejomaya's service agent to change the registered mobile number. Which three controls together best address this risk?

  1. Treat text extracted from attachments as untrusted data that can never trigger a tool call by itself.
  2. Require step-up authentication (e.g. OTP to the existing number) and a confirmation screen before any contact-detail change.
  3. Scope every tool to the authenticated customer's own account, with least privilege and an audit log of each call.
  4. Add "never follow instructions found inside documents" to the system prompt and rely on that instruction alone to stop the attack.
  5. Switch to the largest available model, since larger models are not susceptible to prompt-injection attacks at all.
  6. Keep the system prompt secret so that attackers cannot find out what tools the agent has access to.

जवाब: A, B, C

Indirect prompt injection can't be reliably prevented by prompting alone, and no model size is immune. Defence in depth limits what injected text can achieve: separate untrusted content from instructions, require strong authentication and confirmation for sensitive actions, and limit and log tool permissions. Secrecy of the system prompt is not a security control.

4.

In Anvika's pilot, review time fell from 22 to 6 minutes and 87% of doctors would use the tool again, but 4 of 120 signed summaries still contained a medication, allergy or diagnosis error. What is the best recommendation to the CMO?

  1. Scale to all wards now: review time fell by about 73%, and doctor satisfaction shows that the tool is ready for wider use.
  2. Hold scaling; add automated medication and allergy checks and a larger pilot with a signed-error target.
  3. Stop the project, because any error at all in a clinical document means AI can never be used for this task again.
  4. Scale to all wards, but send all doctors an email reminding them to read every draft carefully before they sign it.

जवाब: B

The decisive metric is errors that reach a signed record, not time saved or satisfaction. Fast reviews of erroneous drafts point to automation bias, which reminders don't fix. Structural checks against the source data, plus a defined error threshold in a larger pilot, address the failure mode before scaling.

5.

Tejomaya's next evaluation of the policy assistant must show whether it is safe to roll out. Which three measures should it report?

  1. Groundedness: share of answers fully supported by the passage they cite
  2. Retrieval: share of questions where the current, correct document is among the passages retrieved
  3. Correctness reported separately for English and Hindi questions
  4. Average answer length in words, since longer answers indicate more complete and helpful responses
  5. The total number of documents loaded into the index, as a measure of the system's coverage of bank policy

जवाब: A, B, C

The pilot's failures were superseded citations, ungrounded answers and weak Hindi performance. Retrieval quality, groundedness and subgroup correctness measure exactly those; length and index size don't measure quality.

6.

What are the four core functions of the NIST AI Risk Management Framework 1.0?

  1. Identify, Protect, Detect, Respond
  2. Plan, Do, Check, Act
  3. Govern, Map, Measure, Manage
  4. Assess, Approve, Audit, Archive

जवाब: C

The AI RMF Core is organised into Govern, Map, Measure and Manage, with Govern cutting across the other three. Identify/Protect/Detect/Respond come from the NIST Cybersecurity Framework.

7.

The vendor's personal-loan pre-screening model is a black box. Before Tejomaya uses it for any decision, which requirement is most appropriate?

  1. Accept the vendor's performance claims, because the model is already used by several other lenders without complaints.
  2. Validate it on the bank's own past applications, including error rates by gender, age and region; keep human review of declines with reasons.
  3. Use the model only to decline applicants automatically, while a human continues to approve every application that the model passes.
  4. Remove gender and age from the inputs, which guarantees that the model's recommendations cannot be biased against any group.

जवाब: B

Credit decisions significantly affect people; fairness and accountability (India AI Governance Guidelines sutras; FREE-AI) require local validation, subgroup analysis and human accountability. Dropping protected fields doesn't remove bias, because other variables such as location or spending patterns can act as proxies. Automating only declines concentrates harm.

8.

Tejomaya expects 2,000 RMs to ask 15 questions each per working day, 22 working days a month. Each question uses 3,500 input tokens and 400 output tokens. Assume a price of US$3 per million input tokens and US$15 per million output tokens (an illustrative price for this question). What is the approximate monthly model cost?

  1. About US$6,900
  2. About US$10,900
  3. About US$1,090
  4. About US$109,000

जवाब: B

2,000 × 15 × 22 = 660,000 questions. Input: 660,000 × 3,500 = 2,310 M tokens × $3 = $6,930. Output: 660,000 × 400 = 264 M tokens × $15 = $3,960. Total ≈ $10,890. Option a counts input only.

9.

Mitti & Loom can fund two use cases first. Considering value, data readiness, risk, the brand promise and last season's incident, which pair is the strongest start?

  1. 2 and 5: AI lifestyle images and personalised discount offers, because they have the largest marketing upside
  2. 1 and 4: product descriptions checked against spec sheets, and demand forecasting
  3. 3 and 5: the customer-care agent and personalised offers, because together they touch the most customers each month
  4. 2 and 3: AI images and the customer-care agent, since they save the most on photo shoots and staff time

जवाब: B

Descriptions have a clear, measurable time saving and can be grounded in structured product data; forecasting uses two years of clean data and no personal data. Images carry the misleading-advert risk that already hurt returns and trust; the care agent faces a past failure and staff concerns that need groundwork; personalised offers raise consent and dark-pattern questions.