AI Engineering Roadmap
November 2026 – April 2027

24-Week Senior AI Engineering Roadmap

Build model-level understanding from gradients to transformers, then apply it to production AI systems. Built on senior backend experience in Laravel, Vue/TypeScript, Node, Kotlin, Docker and Kubernetes.

24
Weeks, each with its own page
6–8
Hours per week
144–192
Total hours
3
Portfolio projects
Weeks completed0 / 24

Goal

Target senior applied AI, AI systems, or AI platform engineering roles, rather than resetting to a junior ML position.

Model internals

Understand the machine

Gradients, neural networks, transformers, training and inference.

Production AI

Ship the system

Retrieval, tool calling, evaluation, serving and operations.

Career

Prove it

Three reviewable projects, design defenses and a senior AI narrative.

Time per week

1.5 h conceptual study · 2 h from-scratch implementation · 2 h applied project work · 1 h testing and documentation · 0.5–1.5 h math and papers.

Overall split

35% model internals · 40% production AI · 25% portfolio and career.

Three portfolio projects

Each project is built in stages across the plan and finishes as a reviewable artifact.

Project 01

Mini GPT from scratch

Autograd, neural network, tokenizer, attention, training and generation.

Project 02

Reliable order processing

Document classification, structured extraction, catalog matching, review workflows and evaluation.

Project 03

Inference and evaluation platform

Open-weight serving, benchmarks, telemetry, deployment and operational controls.

Weekly plan

Open any week for the full guide: steps, where to get each resource, what to build and how to check yourself. Progress is saved in this browser.

First seven days

The on-ramp before Week 1 begins in earnest.

  1. Set up Python, uv, NumPy, PyTorch and pytest.
  2. Review chain rule, gradients, matrices and dot products.
  3. Watch the first part of Karpathy’s micrograd lecture.
  4. Write a Value class with addition and multiplication.
  5. Implement backward propagation through a graph.
  6. Run numerical gradient checks against PyTorch.
  7. Commit the code and write down what a gradient means.

Production architecture to practice

The system you will build and defend in interviews.

Vue / TypeScript frontend
→
Laravel / Kotlin / Node business API
→
Python AI microservice
→
PostgreSQL + pgvector
+
Local or cloud model

Design rules

  • Keep authorization, validation, financial and business operations, and audit controls in deterministic code.
  • Treat model outputs as untrusted input.
  • Require human approval for consequential writes.

Evaluation scorecard

Targets for the projects, and the metrics to track throughout.

98%+

Document classification accuracy

95%+

SKU exact-match

0

Unauthorized model-triggered writes

Pipeline metrics

Field-level precision and recall, false automations, cost per correctly processed document, p95 latency, human-review fraction and regressions.

Serving metrics

Time to first token, tokens per second, p95 latency, concurrent throughput, KV cache and peak memory, cost per successful request.

These are illustrative targets, not measured results.

Career milestones

Checkpoints that show whether the plan is on track.

Week 8

Implement and explain gradients, attention and a transformer.

Week 16

Demonstrate an evaluated, integrated and operational AI service.

Week 24

Present three reviewable projects with architecture, baselines, performance, failure analyses and trade-offs.

Target roles

Senior Applied AI Engineer · Senior AI Systems Engineer · AI Platform Engineer.

Reality check

Experience running real production systems matters more than certificates. Six months will not substitute for years of production ML or research experience.

Technology priorities

PythonNumPyPyTorch Hugging FacePEFTFastAPI pgvectorOllamallama.cpp vLLMDockerKubernetes OpenTelemetryEval tooling

Hardware, budget and confidentiality

Hardware and budget

Your existing laptop is enough for micrograd and mini-GPT experiments. Use Apple Silicon acceleration where available. Rent GPU compute only when a larger model requires it, and cap costs. A provisional $25–$100 per month may cover modest API and occasional GPU use. Training can cost much more.

Confidentiality

Use synthetic or expressly authorized anonymized datasets for portfolio work. Do not publish proprietary source code, credentials, customer documents or company data.

Recommended specialization

AI Systems Engineering with strong model-level understanding. Use your backend and infrastructure strengths to build the dependable system around probabilistic models.

Resources

Links from your plan. Each week page lists the specific resources for that week.

Model internals

Hugging Face and fine-tuning

Building AI applications and evaluation

Serving and inference