Detailed execution plan
12-week AI engineering transition plan
Use this as a practical build schedule. Each week should produce a visible artifact: a repo commit, deployed feature, eval report, design note, or demo video.
Weeks 1-2
LLM fundamentals and API fluency
- Make calls to at least two LLM providers and compare latency, cost, quality, and streaming behavior.
- Build reusable .NET or TypeScript clients with retries, timeouts, structured errors, and token accounting.
- Write notes explaining tokens, context windows, temperature, top-p, hallucination, and tool calling.
Weeks 3-4
Prompt systems and structured outputs
- Create prompt templates for extraction, classification, summarization, and reasoning workflows.
- Add JSON schema validation, refusal handling, prompt versioning, and regression examples.
- Build a small prompt playground that compares model outputs across prompt versions.
Weeks 5-6
Embeddings and RAG
- Implement document ingestion, chunking, embedding, vector storage, retrieval, and source citations.
- Compare chunk size, overlap, metadata filtering, hybrid search, and reranking strategies.
- Deploy a document Q&A chatbot with real docs and a visible evaluation set.
Weeks 7-8
Agents and tool use
- Build a tool-using assistant that can call search, database, calendar, or internal API tools safely.
- Design tool schemas, guardrails, memory boundaries, execution logs, and human approval steps.
- Document agent failure modes and how your app recovers from them.
Weeks 9-10
Evals, observability, and deployment
- Create a golden dataset, automated eval pipeline, and dashboard for quality, latency, and cost.
- Add Langfuse/OpenTelemetry-style tracing for prompts, retrieval, tool calls, and user feedback.
- Containerize the AI service and deploy it with environment-based model routing.
Weeks 11-12
Portfolio, workplace POC, and interviews
- Polish two portfolio projects with README, architecture diagram, screenshots, demo link, and cost notes.
- Prepare an internal AI proposal for your current workplace with risks, ROI, and rollout plan.
- Practice RAG, agent, eval, token-limit, cost-optimization, and system-design interview questions.
Phase 1
LLM foundation and application basics
Mental models, API behavior, prompt design, and production-ready client code
Phase 2
Embeddings, search, and RAG systems
The core skill set behind practical enterprise AI applications
Phase 3
Agents, tools, and workflow orchestration
Build assistants that do useful work while staying observable and controlled
Phase 4
Evaluations, safety, and AI platform engineering
What separates demos from production systems people can trust
Phase 5
Workplace adoption, portfolio, and interviews
Turn learning into visible proof for internal opportunities and AI engineer roles
Portfolio proof
Projects that make the roadmap credible
These are designed for a .NET full-stack developer moving into AI. Each project should include architecture, screenshots, deployment notes, evals, and cost/latency measurements.
Project 1
Document Q&A RAG app
- Upload PDFs/docs, chunk content, generate embeddings, retrieve with citations, and stream answers.
- Add metadata filters, reranking, query rewriting, and a no-answer fallback.
- Publish eval results for faithfulness, answer relevancy, retrieval precision, and latency.
Project 2
AI support triage assistant
- Classify tickets, extract fields, suggest replies, and route to teams using structured outputs.
- Include human approval, audit logs, prompt versioning, and regression tests.
- Show measurable impact: time saved, accuracy, escalation rate, and cost per ticket.
Project 3
Tool-using workflow agent
- Let the agent call safe tools such as search, calendar, CRM mock API, database lookup, or email draft.
- Add tool schemas, permission gates, retries, fallback routes, and trace visualization.
- Document failure cases and how the system prevents unsafe or incorrect actions.
Interview preparation
Questions to rehearse out loud
Prepare concrete answers with project examples. AI engineering interviews reward people who can explain trade-offs, failure modes, and operational details.
RAG and retrieval questions
- How do you choose chunk size and overlap?
- How do you know retrieval is good enough?
- When would you use hybrid search or reranking?
- How do you handle answers that are not present in the documents?
- How do you keep citations accurate?
Agents and production questions
- What can go wrong when an LLM calls tools?
- Where do you put human approval in an agent workflow?
- How do you manage context as an agent runs multiple steps?
- How do you prevent prompt injection in tool-using systems?
- How do you debug a wrong agent decision?
Evals, cost, and architecture questions
- How would you evaluate a summarization or Q&A feature?
- What metrics do you track after launch?
- How do you reduce token cost without hurting quality?
- When would you fine-tune instead of using RAG?
- How would you design an AI assistant for an enterprise app?