GCP AI Upskilling Applied AI for Data Science Trainer

# Track 2: Applied AI for Data Science (Type B) Part of the GCP-Centered AI Upskilling Program. Stevens Institute of Technology, delivered by BetterFutureLabs. 12-week live-virtual cohort starting August 17, 2026. Learner commitment: 8 to 10 hours per week. Teaching days are Tuesday and Thursday. Weeks 1 to 10 are instructor-led builds and labs; Weeks 11 to 12 are the capstone. ## Who this track is for This track is built for the 265 practicing data scientists, applied machine learning practitioners, and analytics professionals at the end-client insurance company who design, validate, and communicate AI/ML and GenAI solutions on Google Cloud. These learners frame problems, prepare data, build and evaluate models, and decide which approach is sound enough to recommend, across insurance work such as claims severity and frequency modeling, underwriting risk, fraud detection, churn, actuarial analysis, and customer-support automation. Their work products are analyses, models, evaluation evidence, and the documentation that hands a validated solution to engineering for productionization. The track focuses on applied model development and enterprise data use rather than shipping the production service, keeping the handoff to engineering clean, complete, and implementation-ready. ## What learners will be able to do by the end - Frame an applied AI or ML problem from a business or operational need, including success criteria and constraints. - Perform feature engineering and data preparation against enterprise data, primarily through BigQuery. - Select and evaluate models, comparing options on performance, cost, and fit. - Design a generative AI solution, including a RAG and grounded AI workflow that ties output to approved data. - Evaluate a GenAI or RAG solution with a gold question-and-answer set, measuring retrieval precision and recall, groundedness and faithfulness, and answer relevance, and logging hallucinations as named evaluation evidence. - Design experiments and validation plans that avoid leakage, use baselines and cross-validation, and track experiments so results are reproducible and statistically defensible. - Evaluate bias, risk, and performance against insurance-regulatory fairness expectations (proxy and protected-class proxies, disparate impact, the NAIC AI bulletin and Department of Insurance review), and document limitations honestly. - Produce model documentation and explainability evidence a reviewer can trust. - Translate data science work into implementation requirements and architecture inputs for engineering. - Assemble a software-engineering-ready handoff package and present design choices and business value to a cross-functional audience. ## Primary skill areas - Applied AI and ML problem framing - Feature engineering and data preparation - Model selection and evaluation - Generative AI solution design - RAG and grounded AI workflows - GenAI and RAG evaluation (retrieval quality, groundedness, hallucination detection) - Statistical validity and leakage-safe practice - Experiment design and validation - Bias, risk, and performance evaluation - Model documentation and explainability - Translation of data science work into implementation requirements - Cross-functional handoff to software engineering and platform teams ## Google Cloud emphasis - Vertex AI and Gemini model workflows - BigQuery for analytics and enterprise data access - Vertex AI notebooks or equivalent development environments - Model evaluation, GenAI and RAG evaluation, and experiment tracking - RAG and grounding workflows - BigQuery ML where appropriate - Cloud Storage for data assets - IAM and data access controls - Monitoring and governance for deployed AI systems ## How this track sits on the shared GCP architecture spine All four tracks are built on one shared GCP architecture spine and common foundation, so every learner shares the same vocabulary for cloud projects, identity, AI workflows, retrieval, deployment, observability, and governance. That foundation covers GCP orientation; AI and GenAI foundations for enterprise use; responsible AI, risk, and governance; data access, privacy, and security; RAG and grounding; AI system evaluation and validation; deployment and observability; and capstone architecture and defense. The data science track leans hardest on the AI development, evaluation, and data-access portions of that spine: Vertex AI and Gemini workflows; RAG and grounding; cloud storage, data, and compute foundations (with BigQuery as the primary data surface); security, governance, and responsible AI controls applied to data access and model risk; Cloud Logging, Monitoring, and operational observability; and capstone architecture and final defense. ## Capstone artifact In Weeks 11 to 12, each learner delivers a validated AI/ML or GenAI solution with a software-engineering-ready handoff package, proving they can move from a business problem to a defensible solution and a handoff an engineering team could implement without guesswork. It is not a from-scratch build: each learner selects their capstone problem and insurance data by the end of Week 3, the program builds two pipelines on that data for breadth (predictive in Weeks 4 to 6 and 9, generative AI and RAG in Weeks 7 to 8), most weekly labs from Week 4 onward operate on the learner's own problem and data, and the implementation-requirements artifact accretes across Weeks 4 to 10. Because each learner's capstone is one pipeline, the other's labs run on the shared insurance scenario and the capstone deepens the chosen pipeline to defensible depth. Weeks 11 to 12 are therefore integration, validation, and defense of a solution built all program, not the first build. Required components: - Architecture diagram showing the data, the model or GenAI workflow, and the GCP services involved (first drafted in Week 10 and refined here) - GCP services used, with a short rationale for each choice - Role-specific implementation artifact: the business or operational problem definition, the data and feature strategy, and the model or GenAI workflow design - Risk and validation evidence: validation and evaluation results, an insurance-regulatory fairness analysis (proxy and protected-class proxies, disparate impact in rating and underwriting, the NAIC Model Bulletin on the use of AI by insurers, and state Department of Insurance review and adverse-action expectations) operationalized through a fairness-metric taxonomy (demographic parity, equalized odds, calibration) with at least one mitigation and the tradeoff stated, an actuarial metric (loss-ratio lift or a Gini/Lorenz curve) where the capstone is a frequency or severity model, plus a risk and limitation analysis - GenAI evaluation evidence where the solution is generative: gold-set results, retrieval precision and recall, groundedness and faithfulness scores, answer relevance, and a hallucination log, with any LLM-as-judge scores validated against a human-labeled subset, parallel to the predictive evaluation plan - A plain-language uncertainty statement: confidence, failure modes, and when not to trust the solution - Security and governance considerations, including data access controls, retrieval-time PII and access enforcement, and responsible AI handling - Operational readiness discussion: the consolidated implementation-requirements artifact (data contract, feature specs, model interface and serving mode, versioning and rollback, monitoring and drift triggers, fairness constraints, and insurance-regulatory flags such as adverse-action and model-governance obligations and NAIC AI bulletin and DOI review points) plus handoff documentation for engineering - Final presentation and defense explaining design choices and business value ## How learners are evaluated Evaluation uses the program-wide rubric, shared across all four tracks: Applied Execution and Capstone Evidence 30%, Technical and AI Capability 25%, Cloud/GCP and Architecture Readiness 20%, Judgment/Validation/Critical Thinking 15%, and Engagement/Adoption/Professional Growth 10%. Evidence is drawn from weekly assignments, hands-on GCP labs, the capstone artifact, technical documentation, and the final defense, supplemented by SME evaluation, course assistant observations, peer and team feedback where appropriate, and a post-program applied skills assessment. Per-learner uplift is measured against these same dimensions from a graded Week-Zero applied diagnostic (see program-operations.md section 1), so the post-program assessment, capstone, and readiness scorecards read as movement from a known baseline. Weekly reflections include a lightweight peer review of one teammate's artifact, and the capstone rubric explicitly rewards communicating uncertainty in plain language: stated confidence, named failure modes, and when not to trust the solution. ## Alignment with Google's recommended learning path This track maps to Google's recommended learning for the data scientist role. Google publishes a distinct [Data Scientist Learning Path](https://www.skills.google/paths/504), the primary role match covering both predictive ML and generative AI, and we pair it with [Integrate Generative AI Into Your Data Workflow](https://www.skills.google/paths/1281), the generative-AI-for-data sub-path built around skill badges. Google offers no certification named "Data Scientist"; the closest credential is the [Professional Machine Learning Engineer Certification](https://www.skills.google/paths/17), an optional stretch outcome. The sections below show where our build maps onto Google's path, the gaps we own, and how the certification fits. ### Skills map: client requirements vs. Google's path The "Covered by Google's published path?" column rates only what Google's catalog delivers on its own. The "Where we cover it" column lists the module actually assigned in this track. | Client-required skill area | Covered by Google's published path? | Where we cover it (module assigned in this track) | | --- | --- | --- | | Applied AI and ML problem framing | Partial (introductory framing only) | Introduction to AI and Machine Learning on Google Cloud (W1); BigQuery for Machine Learning (W4); trainer-built Problem Frame One-Pager (W4) | | Feature engineering and data preparation | Partial | Engineer Data for Predictive Modeling with BigQuery ML (W4); Gemini for Data Scientists and Analysts (W5); trainer-built leakage-safe feature pipeline (W4) | | Model selection and evaluation | Partial | Create ML Models with BigQuery ML (W5 to W6); trainer-built frequency-severity GLM and actuarial-metric overlay (W5 to W6, W9); MLOps with Vertex AI: Model Evaluation (W6); Build and Deploy ML Solutions on Vertex AI (W9) | | Generative AI solution design | Partial | Introduction to Vertex AI Studio (W7); Prompt Design in Agent Platform (W7) | | RAG and grounded AI workflows | Partial | Vector Search and Embeddings (W8); trainer-built RAG pipeline (W8) | | GenAI and RAG evaluation | None to Partial | MLOps for Generative AI (W8, MLOps framing only); trainer-built gold-set evaluation, retrieval precision and recall, faithfulness, and human-validated LLM-as-judge (W8) | | Experiment design and validation | Partial | MLOps with Vertex AI: Model Evaluation (W6); Build and Deploy ML Solutions on Vertex AI (W9); trainer-built experiment tracking from W5 and experiment design in W6 | | Statistical validity and leakage-safe practice | None to Partial | Trainer-built: leakage-safe splits (W4), baseline, cross-validation, and significance check (W6) | | Bias, risk, and performance evaluation | Partial | Responsible AI: Applying AI Principles (W10); trainer-built subgroup slice (W6) and insurance-regulatory fairness with metric taxonomy and mitigation (W10) | | Model documentation and explainability | Partial | Build and Deploy ML Solutions on Vertex AI explainability module (W9); trainer-built model card and uncertainty statement (W9) | | Translation of data science work into implementation requirements | None | No explicit Google course; trainer-built, accreted across W4 to W10 | | Cross-functional handoff to software engineering and platform teams | Partial | MLOps: Getting Started (W10, technical seam only, not the handoff and communication discipline); trainer-built handoff package accreted W4 to W10 | ### Gaps we deliberately cover Where Google's path is partial we extend it; where it is absent we own the gap. Each owned gap is trainer content on the learner's own capstone problem, not a one-time touchpoint: - Translation of data science work into implementation requirements. Google has no explicit course. We accrete an implementation-requirements artifact from Week 4 onward (data contract in Week 4, model interface and serving mode in Weeks 6 and 9, monitoring and drift triggers in Week 9, fairness constraints in Week 10) and consolidate it in Week 10 (Thursday) and the capstone. - Cross-functional handoff as a discipline (documentation-for-handoff and stakeholder communication, not just the technical seam). We provide a handoff-package template, consolidate the accreted artifact in Week 10 (Thursday), draft a first architecture diagram there so the capstone diagram is a refinement, and stress-test the package with a mock engineering reviewer. - GenAI and RAG evaluation. Google's data path builds GenAI and RAG but does not teach how to prove they work. We own a dedicated Week 8 (Thursday) cycle: a small gold question-and-answer set, retrieval precision and recall at k and hit rate, faithfulness, groundedness, and answer relevance scored with a rubric and a human-validated LLM-as-judge, and a hallucination log, delivered as named GenAI evaluation evidence parallel to the predictive plan. - Statistical-validity discipline (leakage avoidance, baselines, cross-validation, significance). We teach leakage-safe practice from the first feature lab (Week 4: split before transform, time-aware splits, a "no leakage" checklist item), then add a baseline, cross-validation, and a confidence-interval or significance check with a stated decision rule in Week 6. - Dedicated bias and fairness evaluation, framed as insurance-regulatory fairness rather than generic ML fairness (which Google embeds inside Responsible AI modules rather than a standalone module). We introduce a subgroup fairness slice on the Week 6 plan, then deepen Week 10 (Tuesday) into insurance-regulatory fairness: proxy and protected-class proxies, disparate impact in rating and underwriting, the NAIC Model Bulletin on the use of AI by insurers, state Department of Insurance review and adverse-action expectations, and model-governance obligations, operationalized through a fairness-metric taxonomy (demographic parity, equalized odds, calibration) with at least one mitigation and the tradeoff, building on the Week 2 checklist. This overlay is the program-wide condition in program-operations.md section 9. - Insurance actuarial modeling methods, which Google's path does not address (it teaches generic classification and regression). Because this audience does claims frequency and severity modeling, we add a frequency-severity treatment using a generalized linear model with an appropriate distribution (BigQuery ML supports several GLM families, and XGBoost supports a Tweedie objective) across the modeling weeks (Weeks 5 to 6 and 9), and add at least one actuarial metric (loss-ratio lift or a Gini/Lorenz curve) to the Week 6 plan. - GenAI-specific safety and security, which Google's data path does not cover. We enforce retrieval-time PII and access controls and output guardrails in Week 8, and extend the Week 2 responsible-AI checklist to a GenAI variant (prompt injection, jailbreak, RAG data leakage, what to flag in handoff) in Week 10. - Model documentation and explainability as a standalone competency (model cards, decision logs, and a plain-language uncertainty statement), which Google folds into its deployment and MLOps courses. We make it a named Week 9 (Thursday) deliverable. - Rigorous experiment design and validation, separate from experiment tracking. We introduce lightweight tracking at the first modeling cycle (Week 5) and require it thereafter, teach explicit experiment design (hypothesis, controlled comparison, confounds) in Week 6, and run a tracked scikit-learn or XGBoost cross-validated build in Week 9 (Tuesday). - Deep RAG and grounding. In Google's data path RAG coverage stops at multimodal RAG, while embeddings and Vector Search sit in the developer GenAI path. We give both cycles of Week 8 to RAG: Tuesday builds the embeddings and Vector Search retrieval pipeline, Thursday grounds and evaluates it. Multimodal document extraction moves to an optional office-hours extension. ### Certification as an optional outcome The [Professional Machine Learning Engineer Certification](https://www.skills.google/paths/17) is an optional stretch credential, not a program requirement. Several weeks overlap its content directly: Week 6 (MLOps model evaluation), Week 9 (build, train, track, and explain on Vertex AI; deployment conceptual only, since this track stops before productionization), and Week 10 (MLOps and the path to deployment, monitoring, and governance, treated conceptually). A learner who wants the credential will find our Vertex AI and MLOps weeks cover a meaningful share of the exam's applied ground; the remaining exam-specific topics (for example full production pipeline automation and live deployment) sit outside this track's scope and need separate preparation. ## Week-by-Week Curriculum Each week runs two learning cycles, one anchored on Tuesday and one on Thursday. The only fixed event in each cycle is the live applied workshop the trainer delivers; it is recorded, so a learner with a conflict can catch up. Each cycle follows a fixed three-part order: 1. Foundational pre-work, before the workshop: the assigned Google Cloud Skills Boost sub-module (a single module slice, typically 15 to 60 minutes). Flexible-before; supplies the shared platform vocabulary the workshop assumes. 2. The live applied workshop (fixed Tuesday or Thursday, recorded): the trainer applies the topic to the learner's real work at the end-client insurance company, using scenarios such as claims, underwriting, fraud, and churn. 3. Post-work, after the workshop: the trainer-built lab, where the learner builds or analyzes something for a realistic insurance problem and encodes the skill. Flexible-after, not pinned to Tuesday or Thursday. Thursday builds directly on Tuesday, so each week stands as one self-contained component. Components stack across Weeks 1 to 10; by Week 10 learners hold the full role-readiness stack for an applied AI data scientist on Google Cloud. The shared cloud, AI, and responsible AI foundation is front-loaded in Weeks 1 to 2. Capstone spine: by the end of Week 3 each learner selects a capstone problem and the enterprise insurance data they will use. From Week 4 onward most weekly labs operate on that same problem and data, and the implementation-requirements artifact accretes week over week. The program builds both pipelines on that data for breadth (predictive in Weeks 4 to 6 and 9, generative AI and RAG in Weeks 7 to 8); where a learner's capstone is one pipeline, the other's labs run on the shared insurance scenario. The governed, de-identified insurance data and validated lab assets these labs depend on from Week 3 onward are provisioned as a pre-program workstream (see program-operations.md section 3). Weeks 11 to 12 carry no new Skills Boost assignment. Every link opens the resource's Google Skills catalog page; labs launch from there and require the program's provided Skills Boost access. If a lab link shows a QL401 access-denied error, open it from the catalog page. The weekly time budget follows a per-cycle pattern: a Skills Boost sub-module (15 to 60 minutes), the live workshop (1 hour), and a trainer-built lab (about 2 to 2.5 hours), plus a 30 minute reflection with a short peer review. Most weeks land near 8 hours; the heavier Weeks 8 and 9 run closer to 9, and no week exceeds the 10 hour ceiling. Office hours are not counted. Platform note: items are hosted at skills.google (the current Google Skills brand); the older cloudskillsboost.google links resolve to the same pages. ### Week 1: Google Cloud foundations, the console, and IAM data access **Estimated learner time this week: about 8 hours** (two cycles plus a 30 minute reflection; office hours not counted). **Outcome (pitch to sponsor):** Your team can stand up a governed Google Cloud project and grant least-privilege, scoped read access to enterprise insurance data across BigQuery, Vertex AI, and Cloud Storage. | Day | Assigned Skills Boost sub-module (pre-work) | Applied workshop focus (fixed, live, recorded) | Trainer-built lab (post-work) | | --- | --- | --- | --- | | Tuesday | [Google Cloud Fundamentals: Core Infrastructure](https://www.skills.google/course_templates/60) (projects, the Cloud console, and IAM access) (about 45 min). | Navigating projects and the console, then mapping IAM roles to least-privilege access on governed enterprise data: who can read which dataset (claims, policy, actuarial). (1 hr, live) | Lab "Project, Console, and IAM Setup": stand up a project, locate the BigQuery, Vertex AI, and Cloud Storage surfaces, and grant scoped read access to a sample insurance dataset. Entry point, no prior lab. (about 2 hr) | | Thursday | [Introduction to AI and Machine Learning on Google Cloud](https://www.skills.google/course_templates/593) (the data-to-AI lifecycle, where Vertex AI and BigQuery ML fit) (about 45 min). | Framing where predictive ML, BigQuery ML, and Gemini or Vertex AI each fit in a data scientist's workflow, and which problems suit which tool. (1 hr, live) | Lab "Data-to-AI Tour": trace one sample insurance problem (for example a claims dataset) across BigQuery, Vertex AI, and Cloud Storage to see the lifecycle end to end. Builds on Tuesday's project and access. (about 2 hr) | ### Week 2: Generative AI and responsible AI foundations **Estimated learner time this week: about 8 hours** (two cycles plus a 30 minute reflection; office hours not counted). **Outcome (pitch to sponsor):** Your team can decide when an insurance business request calls for generative AI versus predictive ML and frame each build inside a responsible AI checklist covering data sourcing, intended use, and known limits, a checklist this track returns to and deepens into fairness work in Week 10. | Day | Assigned Skills Boost sub-module (pre-work) | Applied workshop focus (fixed, live, recorded) | Trainer-built lab (post-work) | | --- | --- | --- | --- | | Tuesday | [Introduction to Generative AI](https://www.skills.google/course_templates/536) (about 45 min): what generative AI is, how it is used, and how it differs from predictive ML. | A live triage clinic, not a lecture: working from real insurance business requests (claims summarization, fraud scoring, churn prediction, policy Q&A), learners decide generative versus predictive for each and defend the call. (1 hr, live) | Lab "GenAI vs Predictive Triage": sort the same slate of real insurance business requests into generative versus predictive approaches with a short rationale for each. Builds on the Week 1 lifecycle map. (about 2 hr) | | Thursday | [Introduction to Responsible AI](https://www.skills.google/course_templates/554) (about 15 min) plus a short supplementary responsible-AI reading from the trainer (about 20 min). | Google's AI principles applied to the data-access, fairness, and disclosure decisions a data scientist actually makes on regulated insurance data. (1 hr, live) | Lab "Responsible AI Checklist v1": draft the project's first responsible AI checklist covering data sourcing, intended use, and known limits; reused as a GenAI safety variant in Week 8 and deepened into a fairness-metric taxonomy in Week 10. Builds on Tuesday's GenAI versus predictive framing. (about 2 hr) | ### Week 3: BigQuery for enterprise data access, analytics, and capstone selection **Estimated learner time this week: about 8 hours** (two cycles plus a 30 minute reflection; office hours not counted). **Outcome (pitch to sponsor):** Your team can query governed enterprise insurance data in BigQuery with attention to cost and access scope, turn the results into a defensible, query-backed insight, and each learner has selected the capstone problem and data they will carry through the rest of the program. | Day | Assigned Skills Boost sub-module (pre-work) | Applied workshop focus (fixed, live, recorded) | Trainer-built lab (post-work) | | --- | --- | --- | --- | | Tuesday | [BigQuery for Data Analysts](https://www.skills.google/course_templates/865) (ingest, transform, and query) (about 45 min). | Querying enterprise insurance tables (claims, policy, customer) to answer a real analytic question, with attention to query cost and data-access scope. (1 hr, live) | Lab "First Enterprise Query": write SQL against a sample insurance dataset and summarize one defensible insight. Builds on the Week 1 IAM data access. (about 2 hr) | | Thursday | [Derive Insights from BigQuery Data](https://www.skills.google/course_templates/623) (about 45 min): foundational SQL querying and building a report from BigQuery data. | Turning raw query results into a defensible insight, and scoping a capstone problem against the insurance data actually available. (1 hr, live) | Lab "Insight Report and Capstone Selection": build a small query-backed report from candidate insurance data, then select your capstone problem (for example claims severity or frequency modeling, underwriting risk, fraud detection, or churn) and the dataset you will use in every lab from Week 4 onward. This selection starts the capstone spine. Builds on Tuesday's querying. (about 2 hr) | ### Week 4: Problem framing and feature engineering **Estimated learner time this week: about 8 hours** (two cycles, with a heavier leakage-safe pipeline lab balanced by a short 15 minute pre-work module, plus a 30 minute reflection; office hours not counted). **Outcome (pitch to sponsor):** Your team can frame an insurance business need as an ML problem with measurable success criteria, then build a repeatable, leakage-safe feature pipeline that produces a clean, model-ready feature table, with a data contract and feature spec captured for the eventual engineering handoff. | Day | Assigned Skills Boost sub-module (pre-work) | Applied workshop focus (fixed, live, recorded) | Trainer-built lab (post-work) | | --- | --- | --- | --- | | Tuesday | [BigQuery for Machine Learning](https://www.skills.google/course_templates/680) (one module, about 45 min). | Framing an insurance business need as an ML problem (target, features, success metric, constraints) and choosing a model type, before any features are built. (1 hr, live) | Lab "Problem Frame One-Pager" (capstone entry deliverable): write the problem definition and measurable success criteria for your capstone problem, then start the implementation-requirements artifact with a data contract (source tables, fields, types, freshness). Builds on the Week 3 capstone selection. (about 2 hr) | | Thursday | [Engineer Data for Predictive Modeling with BigQuery ML](https://www.skills.google/course_templates/627) (about 15 min of this 30 min course). The course also uses Cloud Storage, Dataflow, and Dataprep for ETL; we draw its BigQuery-native transformation portion, since the workshop keeps the feature pipeline in SQL and scheduled queries. Leakage avoidance is the trainer overlay. | Designing a BigQuery-native transformation and feature pipeline (SQL and scheduled queries, no Dataflow) for the framed problem, with leakage avoidance built in from the start. (1 hr, live) | Lab "Leakage-Safe Feature Pipeline": build a BigQuery-native transformation that produces a clean feature table from raw insurance source data; split before you transform, use time-aware splits so no future or target information leaks, and add "no leakage" as an explicit checklist item. Capture feature definitions and online versus offline (serving versus batch) consistency using a minimal Vertex AI Feature Store concept, and add the feature spec to the implementation-requirements artifact. Builds on Tuesday's problem frame. (about 2.5 hr) | ### Week 5: Feature exploration and first BigQuery ML model **Estimated learner time this week: about 8 hours** (two cycles, with a heavier first-model lab, plus a 30 minute reflection; office hours not counted). **Outcome (pitch to sponsor):** Your team can use Gemini to propose and validate candidate features, then train and log a first BigQuery ML model (a GLM with a fitting distribution for claim frequency or severity problems) against the framed capstone problem so results are reproducible. | Day | Assigned Skills Boost sub-module (pre-work) | Applied workshop focus (fixed, live, recorded) | Trainer-built lab (post-work) | | --- | --- | --- | --- | | Tuesday | [Gemini for Data Scientists and Analysts](https://www.skills.google/course_templates/879) (one module, about 45 min). | Using Gemini inside BigQuery to accelerate exploration, feature ideas, and data documentation without losing rigor. (1 hr, live) | Lab "Gemini-Assisted Feature Exploration": use Gemini to profile your capstone dataset and propose candidate features, then validate the proposals by hand. Builds on the Week 4 leakage-safe feature pipeline. (about 2 hr) | | Thursday | [Create ML Models with BigQuery ML](https://www.skills.google/course_templates/626) (Getting Started with BigQuery ML lab) (about 15 min). | Training and reading a first BigQuery ML model against the framed capstone problem, choosing a model type that fits the target (a classifier or linear model, or a GLM with an appropriate distribution, for example Poisson for claim frequency or a gamma or Tweedie family for claim severity, where the capstone is a frequency or severity problem), and starting a lightweight experiment-tracking habit. (1 hr, live) | Lab "First BQML Model with Tracking": train a model in BigQuery ML on your capstone problem, using a GLM with a distribution that matches the target where that fits, interpret its baseline output against the success criteria, and start a lightweight experiment-tracking log (parameters, data version, metrics) that you keep for every later run. Builds on Tuesday's feature exploration. (about 2.5 hr) | ### Week 6: Model selection, evaluation, and statistically valid validation **Estimated learner time this week: about 8 hours** (a lighter Tuesday cycle plus a heavier Thursday validation-plan cycle, plus a 30 minute reflection; office hours not counted). **Outcome (pitch to sponsor):** Your team can compare candidate models against a baseline with cross-validation, an actuarial metric where the target is a frequency or severity model, and a stated decision rule, examine performance across subgroups, and produce an evaluation and validation plan that states which model is sound enough to recommend and how it holds up statistically. | Day | Assigned Skills Boost sub-module (pre-work) | Applied workshop focus (fixed, live, recorded) | Trainer-built lab (post-work) | | --- | --- | --- | --- | | Tuesday | [Create ML Models with BigQuery ML](https://www.skills.google/course_templates/626) (classification and Bracketology labs) (about 15 min). | Experiment design (hypothesis, controlled comparison, confounds) and comparing model types against a simple baseline, reading the evaluation metrics that matter for the problem at hand. (1 hr, live) | Lab "Model Bake-Off with Baseline": train two model types on your capstone problem against a simple baseline (majority-class or a rules heuristic), for example a GLM with a fitting distribution and a tree-based model where the target is a frequency or severity problem, compare evaluation metrics, and apply a temporal or k-fold resampling check; continue the Week 5 experiment-tracking log. Builds on the Week 5 first BQML model. (about 2 hr) | | Thursday | [Machine Learning Operations (MLOps) with Vertex AI: Model Evaluation](https://www.skills.google/course_templates/1080) (one module, about 45 min). | A metrics clinic, not a lecture: learners bring Tuesday's bake-off results and choose the evaluation metrics that fit each insurance problem (including an actuarial metric where the target is a frequency or severity model), define what counts as a valid result, and state a decision rule before looking at the numbers. (1 hr, live) | Lab "Evaluation and Validation Plan": write an evaluation and validation plan with metrics including at least one actuarial metric where the capstone is a frequency or severity model, a holdout or cross-validation strategy, a confidence-interval or significance check, a "no leakage" verification, a subgroup fairness slice (so fairness is examined well before Week 10), and a stated decision rule (effect size or confidence interval). Add the model interface and serving mode (batch versus online) to the implementation-requirements artifact. Builds on Tuesday's bake-off. (about 2.5 hr) | ### Week 7: Generative AI solution design with Vertex AI Studio and prompt design **Estimated learner time this week: about 8 hours** (two cycles plus a 30 minute reflection; office hours not counted). **Outcome (pitch to sponsor):** Your team can design a generative AI solution in Vertex AI Studio and engineer prompts that return reliable, structured output a downstream insurance system can parse. | Day | Assigned Skills Boost sub-module (pre-work) | Applied workshop focus (fixed, live, recorded) | Trainer-built lab (post-work) | | --- | --- | --- | --- | | Tuesday | [Introduction to Vertex AI Studio](https://www.skills.google/course_templates/552) (one module, about 45 min). | Shaping an insurance GenAI use case in Vertex AI Studio (for example claims-note summarization or policy Q&A), from prompt to a testable prototype. (1 hr, live) | Lab "Studio Prototype": build a first Gemini prompt-to-output prototype for a chosen insurance use case, tied to your capstone problem where it is generative. Builds on the Week 2 GenAI versus predictive framing. (about 2 hr) | | Thursday | [Prompt Design in Agent Platform](https://www.skills.google/course_templates/976) (formerly Prompt Design in Vertex AI) (about 45 min of this 1 hour course). Supplies core prompt-engineering and structured-output techniques; the named patterns (zero-shot, few-shot, chain-of-thought) are developed in the workshop and lab. | Zero-shot, few-shot, and chain-of-thought prompting, plus system instructions for structured, parseable output. (1 hr, live) | Lab "Structured Prompt": engineer a prompt that returns reliable structured (JSON) output for a downstream insurance system (for example extracted claim fields). Optional office-hours extension for scanned claims or policy documents: multimodal document extraction (Inspect Rich Documents with Gemini Multimodality and Multimodal RAG). Builds on Tuesday's Studio prototype. (about 2 hr) | ### Week 8: RAG retrieval and GenAI evaluation **Estimated learner time this week: about 9 hours** (a retrieval-build cycle plus a grounded-generation-and-evaluation cycle, both heavier than standard, plus a 30 minute reflection; office hours not counted). **Outcome (pitch to sponsor):** Your team can build a RAG workflow that uses embeddings and Vector Search to ground answers in approved insurance content with citations, and produce GenAI evaluation evidence (retrieval precision and recall, groundedness and faithfulness, answer relevance, and a hallucination log) that proves answers are actually grounded, not merely citation-decorated. | Day | Assigned Skills Boost sub-module (pre-work) | Applied workshop focus (fixed, live, recorded) | Trainer-built lab (post-work) | | --- | --- | --- | --- | | Tuesday | [Vector Search and Embeddings](https://www.skills.google/course_templates/939) (one module: embeddings and Vertex AI Vector Search for retrieval) (about 45 min). | Retrieval mechanics for insurance content: chunking strategy, embeddings, indexing, top-k retrieval, and enforcing retrieval-time access controls and PII handling so retrieval cannot leak restricted policyholder data. (1 hr, live) | Lab "Vector-Search Retrieval": chunk approved insurance content (for example policy wordings or underwriting guidelines), generate embeddings, index them in Vertex AI Vector Search, retrieve the top chunks, and tune chunk size and top-k while enforcing retrieval-time PII and access controls. The Vector Search index and seed insurance corpus are pre-staged before the lab (see program-operations.md section 4), so the lab is spent on chunking and retrieval tuning, not index setup. Builds on the Week 7 structured prompting. (about 2.5 hr) | | Thursday | [Machine Learning Operations (MLOps) for Generative AI](https://www.skills.google/course_templates/927) (about 30 min). Supplies MLOps framing for deploying and managing generative AI on Vertex AI; the GenAI-evaluation vocabulary (groundedness, faithfulness, hallucination, LLM-as-judge) is taught in the workshop and lab, not this course. | Grounded generation and how to evaluate a GenAI or RAG system: groundedness and faithfulness, retrieval precision and recall, answer relevance, hallucination detection, and LLM-as-judge, plus output guardrails. (1 hr, live) | Lab "RAG Evaluation Evidence": ground answers in the retrieved chunks with citations, then build a small gold question-and-answer set; measure retrieval precision and recall at k and hit rate; score faithfulness, groundedness, and answer relevance with a rubric and an LLM-as-judge, but first validate the judge against a small human-labeled subset and report judge-to-human agreement so its scores are trusted only where it matches human labels; log hallucinations; and add output guardrails. Produce "GenAI evaluation evidence" as a named deliverable parallel to the predictive evaluation plan. Builds on Tuesday's retrieval pipeline. (about 2.5 hr) | ### Week 9: Python notebook build, explainability, and documentation **Estimated learner time this week: about 9 hours** (a Python-build cycle plus an explainability-and-documentation cycle, both heavier than standard, plus a 30 minute reflection; office hours not counted). **Outcome (pitch to sponsor):** Your team can build a tracked scikit-learn or XGBoost model with cross-validation and basic tuning in Vertex AI Workbench, compare it against the BigQuery ML baseline, and produce feature-attribution explainability evidence, a model card, and a plain-language uncertainty statement a reviewer can trust. | Day | Assigned Skills Boost sub-module (pre-work) | Applied workshop focus (fixed, live, recorded) | Trainer-built lab (post-work) | | --- | --- | --- | --- | | Tuesday | [Build and Deploy Machine Learning Solutions on Vertex AI](https://www.skills.google/course_templates/684) (one lab module: Vertex AI Workbench notebook, train and track) (about 45 min). | Moving from SQL-based BigQuery ML to a Python notebook build in Vertex AI Workbench: scikit-learn or XGBoost with k-fold cross-validation and basic hyperparameter tuning (including an XGBoost Tweedie objective for frequency or severity targets), compared against the BigQuery ML baseline. (1 hr, live) | Lab "Python Notebook Build (scikit-learn or XGBoost)": in a Vertex AI Workbench notebook, build a scikit-learn or XGBoost model on your capstone problem with k-fold cross-validation and basic tuning (using an XGBoost Tweedie objective for frequency or severity targets), compare it against the Week 5 and Week 6 BigQuery ML baseline on the same metrics including the actuarial metric where it applies, and continue the experiment-tracking log. The Workbench environment and starter scaffold are pre-staged before the lab (see program-operations.md section 4). Builds on the Week 6 evaluation and validation plan. (about 2.5 hr) | | Thursday | [Build and Deploy Machine Learning Solutions on Vertex AI](https://www.skills.google/course_templates/684) (one lab module: evaluate and explain; deployment treated conceptually since this track stops before productionization) (about 45 min). | Generating explainability evidence, writing model documentation that survives review, and communicating uncertainty in plain language. (1 hr, live) | Lab "Explainability, Model Card, and Uncertainty": produce feature-attribution evidence, a short model card, and a plain-language uncertainty statement (confidence, failure modes, when not to trust the model). Add model versioning and rollback and monitoring and drift triggers to the implementation-requirements artifact. Builds on Tuesday's Python build. (about 2.5 hr) | ### Week 10: Fairness in depth and the engineering handoff **Estimated learner time this week: about 8 hours** (a fairness cycle plus a handoff-consolidation cycle plus a 30 minute reflection; office hours not counted). **Outcome (pitch to sponsor):** Your team can evaluate an insurance solution for bias and fairness against insurance-regulatory expectations (the proxies, disparate impact, NAIC Model Bulletin, and Department of Insurance review and adverse-action expectations above) using a fairness-metric taxonomy with at least one mitigation, and consolidate the validated work, including its insurance-regulatory flags, into a clean, implementation-ready handoff package with a first architecture diagram that an engineering team could pick up without guesswork. | Day | Assigned Skills Boost sub-module (pre-work) | Applied workshop focus (fixed, live, recorded) | Trainer-built lab (post-work) | | --- | --- | --- | --- | | Tuesday | [Responsible AI: Applying AI Principles with Google Cloud](https://www.skills.google/course_templates/388) (one module, about 45 min). The insurance-regulatory specifics (proxy and protected-class proxies, disparate impact, the NAIC AI bulletin and Department of Insurance expectations) are the trainer overlay (see program-operations.md section 9). | A fairness clinic on the learner's own model, not a lecture: learners bring the model and subgroup slice from Weeks 6 and 9 and operationalize insurance-regulatory fairness live, testing for proxy and protected-class proxies and disparate impact in rating and underwriting, applying the fairness-metric taxonomy (demographic parity, equalized odds, calibration), applying a mitigation, and stating the accuracy-versus-fairness tradeoff against NAIC Model Bulletin and Department of Insurance review and adverse-action expectations, plus the GenAI safety risks (prompt injection, jailbreak, RAG data leakage) a generative solution must handle. (1 hr, live) | Lab "Insurance-Regulatory Fairness, Mitigation, and GenAI Risk Checklist": evaluate your model across subgroups using the fairness-metric taxonomy, test for those proxies and disparate impact, apply at least one mitigation and state the tradeoff, and document the insurance-regulatory flags the engineering team must carry (adverse-action obligations, model-governance and documentation expectations, NAIC AI bulletin and DOI review points). Extend the Week 2 responsible-AI checklist into a GenAI variant (prompt injection, jailbreak, retrieval-time PII and access, output guardrails, what to flag in the handoff). Add fairness constraints and the insurance-regulatory flags to the implementation-requirements artifact. Builds on the Weeks 6, 8, 9, and 2 artifacts. (about 2 hr) | | Thursday | [Machine Learning Operations (MLOps): Getting Started](https://www.skills.google/course_templates/158) (one module, about 45 min; teaches the technical seam, not the handoff and communication discipline, which is trainer-built). | Translating data science work into implementation requirements and a clean handoff to software engineering and platform teams (deployment, monitoring, and governance treated conceptually), then drafting a first architecture diagram of the data, the model or GenAI workflow, and the GCP services from the list the handoff already assembles, using a provided template. (1 hr, live) | Lab "Handoff Package Consolidation, Architecture Diagram, and Mock Engineering Review": consolidate the implementation-requirements artifact accrued since Week 4 (data contract, feature specs, model interface and serving mode, versioning and rollback, monitoring and drift triggers, fairness constraints, and insurance-regulatory flags) into a handoff package using the template, draft a first architecture diagram of the data, the model or GenAI workflow, and the GCP services involved (a low-cost add, since the service list, interfaces, and serving mode are already assembled) so the capstone diagram is a refinement, not a first draft, then defend the package and diagram to a mock engineering reviewer who stress-tests them for missing schemas, serving mode, latency and throughput, failure modes, and retraining triggers. Builds on every prior week. (about 2 hr) | ### Weeks 11 to 12: Capstone No new Skills Boost assignment. Because most labs from Week 4 onward operated on the learner's own capstone problem and data across both a predictive pipeline (Weeks 4 to 6 and 9) and a generative AI and RAG pipeline (Weeks 7 to 8), these two weeks are integration, validation, and defense, not a first build. Each learner deepens the pipeline their capstone is built on to defensible depth and carries the other as demonstrated breadth. Learners assemble the accumulated stack (problem framing, leakage-safe BigQuery feature work, model selection and statistically valid validation including an actuarial metric where the target is a frequency or severity model, generative AI and grounded RAG design with GenAI evaluation evidence, the cross-validated Python build, insurance-regulatory fairness analysis with mitigation, explainability and uncertainty communication, and the consolidated engineering handoff with its insurance-regulatory flags) into a validated applied AI/ML or GenAI solution, then present and defend their design choices and business value to a cross-functional audience. ### Sources All links were verified by fetching the page and confirming the title; each entry notes the catalog duration shown. All resolve at the current skills.google domain; the older cloudskillsboost.google addresses redirect to the same pages. A learner is assigned only one sub-module or lab from each resource per cycle (typically 15 to 60 minutes), not the full multi-hour badge, so the duration below is the whole resource, not the assigned slice. - [Google Cloud Fundamentals: Core Infrastructure](https://www.skills.google/course_templates/60) (Course, Introductory; full resource about 5 hours) - [Introduction to AI and Machine Learning on Google Cloud](https://www.skills.google/course_templates/593) (Course, Introductory; full resource about 8 hours) - [Introduction to Generative AI](https://www.skills.google/course_templates/536) (Course, Introductory; about 45 minutes) - [Introduction to Responsible AI](https://www.skills.google/course_templates/554) (Course, Introductory; about 15 minutes) - [BigQuery for Data Analysts](https://www.skills.google/course_templates/865) (Course, Intermediate; full resource about 5 hours) - [Derive Insights from BigQuery Data](https://www.skills.google/course_templates/623) (Course, Introductory; about 45 minutes) - [Engineer Data for Predictive Modeling with BigQuery ML](https://www.skills.google/course_templates/627) (Course, Intermediate; about 30 minutes) - [Gemini for Data Scientists and Analysts](https://www.skills.google/course_templates/879) (Course, Introductory; full resource about 2 hours) - [BigQuery for Machine Learning](https://www.skills.google/course_templates/680) (Course, Introductory; full resource about 4 hours) - [Create ML Models with BigQuery ML](https://www.skills.google/course_templates/626) (Course, Intermediate; about 30 minutes) - [Machine Learning Operations (MLOps) with Vertex AI: Model Evaluation](https://www.skills.google/course_templates/1080) (Course, Intermediate; full resource about 2 hours 30 minutes) - [Introduction to Vertex AI Studio](https://www.skills.google/course_templates/552) (Course, Introductory; full resource about 2 hours) - [Prompt Design in Agent Platform](https://www.skills.google/course_templates/976) (Course, Introductory; about 1 hour; the platform now shows this title, previously published as "Prompt Design in Vertex AI") - [Inspect Rich Documents with Gemini Multimodality and Multimodal RAG](https://www.skills.google/course_templates/981) (Skill badge, Intermediate; about 45 minutes; assigned as an optional Week 7 office-hours extension, not a required cycle) - [Vector Search and Embeddings](https://www.skills.google/course_templates/939) (Course, Intermediate; full resource about 4 hours) - [Machine Learning Operations (MLOps) for Generative AI](https://www.skills.google/course_templates/927) (Course, Intermediate; full resource about 30 minutes) - [Build and Deploy Machine Learning Solutions on Vertex AI](https://www.skills.google/course_templates/684) (Skill badge, Intermediate; full badge about 7 to 8 hours, one lab module assigned per cycle. Title confirmed via the challenge-lab sub-URLs course_templates/684/labs/481171 and course_templates/684/labs/556977 and the external catalog listing.) - [Responsible AI: Applying AI Principles with Google Cloud](https://www.skills.google/course_templates/388) (Course, Introductory; full resource about 2 hours) - [Machine Learning Operations (MLOps): Getting Started](https://www.skills.google/course_templates/158) (Course, Intermediate; full resource about 4 hours 30 minutes)