Twelve weeks.
One connected delivery project.

Start with one repeated decision or handoff. Name the person doing it, the information used, the action taken and current problems. Separate measured starting results from assumptions. AI can help when language or messy information slows the work. Use fixed rules for permissions, identity and transactions. Limit the first release to one user, one workflow and one measurable result.

Deliverable: Submit a discovery brief with: user, workflow, evidence, assumptions, business metric, three discovery questions, proposed scope, exclusions, and human approval points.

Applied AI practitioner track: Prioritize the opportunity against two alternatives using value, effort, and risk.

Engineering track: Add a simple data-flow sketch describing inputs, transformations, outputs, and trust boundaries.

Human checkpoint: Discovery and scope

A useful assistant answers from a limited set of sources and points to the exact supporting record. A confident answer without support is a failure. Define the user, allowed sources, output format and behavior for missing or conflicting information. Documents provide data; their instructions must not override application rules. Test manually with a small source pack first. Add document search when volume or updates require it.

Deliverable: Submit your assistant instruction, source mapping, three test questions with actual answers, one failure you found, and your correction. Include a demo link or transcript.

Applied AI practitioner track: Configure the assistant in your existing tool and explain the limits to a simulated account manager.

Engineering track: Implement a server-side endpoint with input validation and environment-based secrets. Explain why API keys never belong in a browser bundle.

Automated concept and evidence checks guide progression.

A tool call requests an action. It does not prove permission or success. Check inputs, verified identity and user permissions outside the model. Handle results with fixed rules. Require approval for customer-facing actions or actions needing extra privileges. Use a request ID, called an idempotency key, to prevent duplicates on retries. Define what happens after timeouts, unavailable services or uncertain results.

Deliverable: Submit tool contracts, an authorization and approval flow, four test cases with results, retry/duplicate handling, and a handoff example. Practitioners may use a configured workflow or detailed simulation; engineers must show what they built.

Applied AI practitioner track: Build the process configuration or a traceable simulation and document engineering handoff requirements.

Engineering track: Include code or a repository excerpt implementing validation and idempotency. Do not submit credentials.

Human checkpoint: Tool choice and integration

A demo shows one successful run. Delivery evidence also covers failures, cost, monitoring and a named operator. Test realistic cases against agreed criteria before changing the prompt. Track wrong answers, unauthorized actions, requests for human help and cost per completed task. State which tests actually ran. Simulations help with practice but do not prove production behavior. Final skill assessment needs human review and a live change request.

Deliverable: Submit your evaluation table, results and remaining failures, cost estimate with assumptions, runbook, portfolio narrative, and demo/repository link. Request practice feedback on this first-release evidence. The final live assessment takes place in week 12.

Applied AI practitioner track: Include adoption plan, stakeholder demonstration, benefits measurement, and engineering acceptance criteria.

Engineering track: Include deployment evidence, logs with secrets removed, a rollback plan, and checks that another person can run again.

Automated concept and evidence checks guide progression.

List each source, its owner, update schedule, sensitivity and allowed users. Give documents stable IDs and versions. Keep these IDs when splitting documents into sections and indexing them for search. Remove outdated material and records whose access has been withdrawn. If sources do not support an answer, say it is unknown. Use fictional records in labs. Source permissions must be enforced by the application, not a prompt.

Deliverable: Submit a source inventory, access matrix, version/update procedure, five source-selection results and a deletion test.

Applied AI practitioner track: Produce a data-owner approval checklist and manually verify the five source-selection cases.

Engineering track: Implement a source-filter function, version field and a test that denies a source outside the caller’s scope.

Automated concept and evidence checks guide progression.

Define success before changing the application. Check search results, whether answers match sources, action results and user experience separately. Include common tasks and serious failures. Write expected outcomes before viewing answers. Use automated checks for permissions, data formats and duplicate prevention. Use human judgment for usefulness and difficult industry questions. Keep separate cases that are not used to improve the system.

Deliverable: Submit the cases, scoring rules, before-and-after results, separate test results and your recommendation on whether to release.

Applied AI practitioner track: Use a spreadsheet evaluation rubric and independently review a sample of answers.

Engineering track: Automate deterministic assertions and produce a machine-readable result file.

Automated concept and evidence checks guide progression.

Start with clear workflow steps before adding an agent that chooses its next action. Define each state, allowed transition and stopping point. Give tools limited permissions and defined result formats. A model can propose an action; the application must authorize it. Limit retries, time and spending. MCP offers a common tool interface, but identity, permissions and tool input checks remain necessary. Show where a person can pause, override or resume work.

Deliverable: Submit a diagram of workflow states, tool input and output formats, retry limits, an activity record and your reason for choosing this workflow.

Applied AI practitioner track: Configure or simulate the same workflow and provide a precise engineering acceptance checklist.

Engineering track: Implement one workflow state machine and expose a narrowly scoped tool contract suitable for MCP integration.

Automated concept and evidence checks guide progression.

Help the user reach a clear result. Confirm their goal and summarize before actions with serious consequences. Do not ask again for information already available with permission. State uncertainty and offer a useful handoff to a person. Voice adds interruptions, delay, transcription errors and consent needs. A text prototype is enough for the core pathway. Voice practice is optional and depends on access to a suitable service. Never clone a voice without permission.

Deliverable: Submit conversation flows, transcripts, handoff packet, accessibility considerations and measured improvements.

Applied AI practitioner track: Review the journey as a customer and specify measurable experience acceptance criteria.

Engineering track: Implement typed conversation events and confirm tool outcomes before claiming completion.

Automated concept and evidence checks guide progression.

Deploy only the smallest workflow you are authorized to release. Separate development and production settings. Keep secrets on the server and give users only necessary access. Add health checks, safe activity logs, release notes and a way to restore the earlier version. Test the deployed environment with fictional users before inviting real ones. A successful build does not prove deployment works. Record the released version and actual results.

Deliverable: Submit deployment evidence or explicitly labeled simulation, release checklist, smoke results and rollback runbook.

Applied AI practitioner track: Produce an evidence pack for the person who will operate it and review the evidence with a technical partner.

Engineering track: Provide a reproducible deployment command and executable smoke checks.

Human checkpoint: Pilot and release readiness

After launch, track completed tasks, response time, failure rates, requests for human help and cost per result. Message counts alone do not show success. Set thresholds for investigation, request limits or shutdown. Name an operator and response process. Investigate incidents using records with sensitive details removed. Find the cause and add a test to prevent it returning. Separate actual measurements from estimates.

Deliverable: Submit metrics definitions, cost worksheet, incident drill and continuous-improvement backlog.

Applied AI practitioner track: Define service ownership, escalation expectations and the business impact of downtime.

Engineering track: Add structured logs, a health check and rate limiting to your implementation.

Automated concept and evidence checks guide progression.

Define the customer, scope, success criteria, dependencies, business terms and handover. Adoption needs workflow owners, communication and training. Explain what the solution replaces, helps with or leaves unchanged. Measure the starting point before claiming value compared with cost. Connect prior expertise to work you can demonstrate. Do not adopt a professional title that your evidence does not support.

Deliverable: Submit scope, acceptance criteria, adoption plan, portfolio narrative and discovery roleplay notes.

Applied AI practitioner track: Demonstrate business discovery, value framing and a precise engineering handoff.

Engineering track: Explain architecture and implementation tradeoffs to a nontechnical stakeholder.

Automated concept and evidence checks guide progression.

Final readiness is a human judgment based on evidence, not an AI score. Present the problem, solution, build or configuration, tests, operating plan and measured value. Explain what you understand and can change yourself. A live change request checks adaptability. A failure exercise checks how you respond. Completion, assessed skill and eligibility for supervised client work are separate decisions. No result guarantees a job or NeoAistriq assignment.

Deliverable: Submit a complete portfolio pack, demonstration transcript or link, test results, limitations, and a request for live human assessment.

Applied AI practitioner track: Lead the customer demonstration, acceptance discussion and adoption handover.

Engineering track: Demonstrate independent debugging and an end-to-end change with tests another person can run again.

Human checkpoint: Demonstrated delivery and impact