Practical decisions need practical tools.

Stress-test a proposed design and produce a decision record.

Act as a senior cloud AI architect. Review the architecture below for security, reliability, operability, cost, and unnecessary complexity. Separate facts from assumptions. Identify the three highest-risk decisions, propose alternatives with trade-offs, and finish with a concise architecture decision record. Ask for missing constraints before making a recommendation.

Architecture: [paste architecture]
Constraints: [users, regions, data sensitivity, budget, latency, recovery targets]

Investigate weak answers before changing the model.

Diagnose this RAG failure as a retrieval engineer. Evaluate ingestion, chunking, metadata, query transformation, filtering, ranking, context assembly, and answer grounding in that order. For each likely cause, state the evidence needed and one controlled experiment. Do not assume the language model is the problem. Return a prioritized test plan with success metrics.

Failure examples: [paste examples]
Current pipeline: [describe pipeline]
Evaluation set: [describe or attach]

Define controls before an agent can take actions.

Review the following agent tool contract for production risk. Classify every action as read-only, reversible write, irreversible write, or external communication. Identify required authentication, authorization, input validation, rate limits, audit events, approval gates, timeout behavior, and rollback options. Return a table of risks and controls, then list the tests required before release.

Tool contract: [paste schema]
Business process: [describe process]

Create representative tests without inventing expected facts.

Design an evaluation dataset for this AI agent. Create representative happy paths, ambiguous requests, missing-data cases, tool failures, adversarial inputs, and policy boundaries. For each case provide the user input, available context, expected behavior, prohibited behavior, and scoring criteria. Mark any expected answer that requires a domain expert to supply ground truth.

Agent purpose: [purpose]
Tools: [tools]
Policies: [policies]

Expose missing workload assumptions before deployment.

Audit this Azure AI cost estimate. List every explicit and hidden assumption about traffic, tokens, storage, indexing, networking, environments, logging, evaluation, availability, and growth. Distinguish fixed and usage-based drivers. Do not invent prices; provide links or fields that must be checked in the official Azure pricing calculator. Return best-case, expected, and stress-case assumption tables.

Estimate: [paste estimate]
Workload: [describe workload]

Turn an operational failure into measurable follow-up work.

Facilitate a blameless postmortem for this AI system incident. Build a verified timeline, separate contributing conditions from root causes, and cover model, retrieval, tools, data, infrastructure, monitoring, and human approval paths. Flag uncertain claims. Produce corrective actions with owner, priority, verification method, and due-date placeholder.

Incident notes: [paste notes]
Telemetry: [paste evidence]

Practice explaining a real project with engineering judgment.

Interview me about the project below for a [target role] position. Ask one question at a time, starting with architecture and moving through trade-offs, failures, security, cost, and outcomes. After each answer, score clarity, technical depth, evidence, and ownership from 1-5. Then give one specific improvement and a stronger example answer that does not invent experience I did not claim.

Project: [describe project]

Convert task language into truthful, measurable evidence.

Rewrite these resume bullets for a [target role]. Preserve factual accuracy and never invent metrics, scale, tools, or ownership. For each bullet, identify the missing evidence, then produce a concise version using action, technical decision, scope, and outcome. Use [metric needed] where evidence is absent.

Current bullets: [paste bullets]
Target job description: [paste description]