watsonx GenAI Engineer study plan

Map the anatomy of a foundation model: tokenization, embeddings, attention, decoding, and context window limits, and state how each shapes prompt behavior

Compare model families available in watsonx (Granite and third-party options) on size, context length, licensing, and task fit, and justify a model choice for two sample use cases

Write and iterate zero-shot, few-shot, and chain-of-thought prompts for a classification task and a summarization task, recording how each change affected output quality

Identify and fix five common prompt failure modes: ambiguous instructions, conflicting constraints, missing output format, hallucinated facts, and prompt injection exposure

Explain the RAG pipeline end to end: ingestion, chunking, embedding, indexing, retrieval, reranking, and grounded generation

Configure a vector index in watsonx and compare dense, sparse, and hybrid retrieval on the same document set

Tune chunk size and overlap and measure the effect on retrieval precision for a fixed question set

Attach citations and guardrail instructions so generated answers stay grounded in retrieved passages and abstain when evidence is missing

Design a multi-step orchestration flow that chains prompt templates, retrieval, tool calls, and output validation

Implement at least two tool or function calls and handle failure, timeout, and retry paths

Add conversation memory and session state to a chat-style application and verify context does not leak across sessions

Apply input and output guardrails against prompt injection, unsafe content, and PII leakage

Describe the watsonx platform architecture: control plane versus runtime services, projects and spaces, model inference endpoints, and identity and access boundaries

Provision a project, assign roles, and connect a data source and a model endpoint using least-privilege access

Trace a single request from client to model inference and back, naming each component it passes through

Compare watsonx.ai, watsonx.data, and watsonx.governance responsibilities and state which service owns which artifact

Distinguish prompt engineering, RAG, fine-tuning, and parameter-efficient tuning by cost, data requirement, and when each is the right answer

Prepare a small instruction dataset in the required JSONL format with train and validation splits and document your labeling rules

Run a tuning job on a small model and compare tuned versus base output on a held-out prompt set

Evaluate tuning results with task-specific metrics and decide whether to keep, adjust, or abandon the tuned model

Enumerate the governance lifecycle: use case intake, risk assessment, model factsheet, evaluation, approval, and ongoing monitoring

Configure fairness, drift, and quality monitors for a deployed model and interpret the resulting metrics

Draft a model factsheet and risk register entry for your week-3 application, naming owners and mitigation controls

Map your solution to a regulatory or internal policy framework and identify the evidence each control requires

Deploy a tuned or prompted model to a serving endpoint and configure autoscaling, rate limits, and versioning

Instrument the application with logging, latency and token-usage tracking, and quality sampling on live traffic

Build a data preparation pipeline that cleans, deduplicates, chunks, and embeds source documents and can be rerun idempotently

Define promotion, rollback, and deprecation criteria for a model version and document the runbook

Re-derive the key decision points across all eight domains from memory and list every gap with a confidence rating

Retake a timed 60-question mixed practice set and review every wrong answer to its root concept

Rehearse explaining your capstone architecture, tuning decision, and governance controls aloud in under ten minutes

Close remaining weak spots with targeted review rather than full re-reading