Free interactive career roadmap

AI/ML Engineer Roadmap

A complete six-month intermediate AI/ML engineering roadmap covering data preparation, statistical thinking, classical machine learning, deep learning, transformers, evaluation, deployment and MLOps. Build one measurable prediction product from notebook baseline to monitored API, then add a carefully evaluated language-model feature. Mostly official documentation is used throughout.

6 months duration24 guided weeks10–12 hours per week

Your progress0 / 24 weeks

Saved on this device

01Month 1 — Data and Mathematical Foundations

02Month 2 — Classical Machine Learning

03Month 3 — Deep Learning and Transformers

04Month 4 — Production ML and MLOps

05Month 5 — Applied LLM Engineering and Evaluation

06Month 6 — Capstone and AI/ML Interview Readiness

Data and Mathematical Foundations

Build a reproducible Python data workflow and the mathematical intuition needed to reason about models.

W01

Python, NumPy and Vectorised ThinkingMove from general Python code to efficient numerical workflows.

10–12 hours per week0/5

Learn: Study arrays, shapes, dtypes, indexing, broadcasting and vectorisation.Practise: Replace loop-based transformations with tested NumPy operations.Build: Create reusable feature utilities for a small labelled ticket dataset.Verify: Check shapes, types, missing values and deterministic output.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

A tested data-loading and vectorised feature module.

W02

pandas and Data QualityTurn messy records into an auditable modelling table.

10–12 hours per week0/5

Learn: Study DataFrames, joins, grouping, missing data, duplicates and categorical types.Practise: Create explicit checks for schema, nulls, ranges and duplicates.Build: Build a repeatable cleaning pipeline and data dictionary.Verify: Re-run from raw data and compare row counts at every step.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

A clean modelling dataset, validation report and data dictionary.

W03

Probability, Statistics and Linear AlgebraUnderstand the assumptions behind metrics and common models without turning the month into pure theory.

10–12 hours per week0/5

Learn: Review distributions, expectation, variance, sampling, confidence, vectors and matrix operations.Practise: Simulate sampling and measurement uncertainty in a notebook.Build: Calculate simple baselines and confidence intervals for ticket outcomes.Verify: Explain correlation versus causation and one sampling bias in the dataset.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

A concise notebook connecting statistical concepts to the project data.

W04

Problem Framing and BaselinesDefine a model only after clarifying the user decision, target and cost of mistakes.

10–12 hours per week0/5

Learn: Study target definition, leakage, train/test boundaries and baseline models.Practise: Write a prediction contract: input, output, metric, threshold and non-goals.Build: Create heuristic and dummy baselines for ticket priority or routing.Verify: Confirm the split prevents future or customer leakage.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

Month 1 baseline report with problem statement, data split and acceptance metric.

MONTH 01 MILESTONEA versioned exploratory analysis and baseline problem definition for support-ticket classification.

Classical Machine Learning

Build, compare and explain reliable supervised-learning pipelines.

W05

Preprocessing and PipelinesMake every transformation reproducible and safe from leakage.

10–12 hours per week0/5

Learn: Study encoding, scaling, imputation, column transforms and pipelines.Practise: Build preprocessing for numeric, categorical and text columns.Build: Create one end-to-end scikit-learn Pipeline from raw row to prediction.Verify: Fit transforms on training data only and add an unseen-category test.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

A serialisable preprocessing-and-model pipeline with leakage tests.

W06

Regression and Classification ModelsUnderstand how simple models differ before reaching for complex algorithms.

10–12 hours per week0/5

Learn: Study linear/logistic models, trees, ensembles, regularisation and bias-variance trade-offs.Practise: Train at least three model families under the same split and metric.Build: Add a model comparison table with training and inference costs.Verify: Explain one case where the simplest model should win.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

A fair comparison of three candidate models with a selected baseline challenger.

W07

Evaluation, Cross-Validation and ThresholdsMatch evaluation to the real cost of false positives and false negatives.

10–12 hours per week0/5

Learn: Study cross-validation, precision, recall, ROC/PR, calibration and threshold selection.Practise: Evaluate by segment and inspect variance across folds.Build: Choose an operating threshold using an explicit cost or service constraint.Verify: Create a confusion-matrix error sample and review misclassified rows.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

An evaluation report with confidence, segment metrics, threshold and error examples.

W08

Feature Engineering and ExplainabilityImprove signal while keeping features valid at prediction time.

10–12 hours per week0/5

Learn: Study feature selection, transformations, text features and model inspection.Practise: Create candidate features from timestamps, text and customer history.Build: Add permutation importance and local explanations for selected cases.Verify: Remove any feature unavailable or unstable at inference time.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

Version 2 model with documented features, explanations and an error-analysis update.

MONTH 02 MILESTONEA cross-validated model that beats the baseline and has an error analysis tied to user impact.

Deep Learning and Transformers

Learn neural-network workflows and use pretrained models responsibly for text.

W09

PyTorch Tensors and Training LoopsUnderstand tensors, autograd, modules, losses and optimisation as one complete loop.

10–12 hours per week0/5

Learn: Study tensors, datasets, dataloaders, autograd and nn.Module.Practise: Train a small network and inspect overfitting on train and validation curves.Build: Create a reproducible PyTorch training script for ticket classification.Verify: Save and reload weights and reproduce an evaluation result.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

A configurable PyTorch training pipeline with saved model and metrics.

W10

Neural Network Design and RegularisationReason about architecture and training behaviour instead of blindly adding layers.

10–12 hours per week0/5

Learn: Study activations, losses, optimisation, dropout, normalisation and early stopping.Practise: Run controlled experiments changing one factor at a time.Build: Add callbacks or code for checkpointing and early stopping.Verify: Compare learning curves and explain underfit versus overfit evidence.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

An experiment note showing how two training choices changed generalisation.

W11

Transformers and Pretrained ModelsUse pretrained language models with clear task, data and compute boundaries.

10–12 hours per week0/5

Learn: Study tokenisation, embeddings, attention, pipelines and fine-tuning workflow.Practise: Benchmark a pretrained text-classification pipeline on held-out tickets.Build: Fine-tune a compact model only if the benchmark and data justify it.Verify: Measure quality, latency and memory against the classical model.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

A classical-versus-transformer benchmark with reproducible inference code.

W12

Experiment Tracking and ReproducibilityMake every meaningful model result traceable to code, data and configuration.

10–12 hours per week0/5

Learn: Study runs, parameters, metrics, artifacts, seeds and environment capture.Practise: Log comparable classical and deep-learning experiments.Build: Add MLflow tracking and a repeatable training command.Verify: Reproduce the selected run from a clean environment.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

Month 3 model registry candidate with tracked experiments and reproducibility checklist.

MONTH 03 MILESTONEA reproducible PyTorch text model and a transformer benchmark with a justified choice.

Production ML and MLOps

Package, serve, test and monitor the model as a production dependency.

W13

Model Packaging and RegistryVersion model code, schema and artifacts together.

10–12 hours per week0/5

Learn: Study model signatures, lineage, registry aliases and promotion workflow.Practise: Register candidate and champion models with inputs and metadata.Build: Package the selected pipeline in MLflow with an explicit signature.Verify: Load by registry alias and verify schema rejection for bad requests.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

A registered model with lineage, signature, example input and promotion notes.

W14

Inference API and ContainerServe predictions with validation, latency awareness and safe error handling.

10–12 hours per week0/5

Learn: Study FastAPI schemas, model loading, batch inference and container builds.Practise: Load-test single and batch prediction endpoints.Build: Build an API that returns prediction, confidence and model version.Verify: Containerise and verify startup, health and invalid-input behaviour.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

A Dockerised inference API with versioned responses and load-test results.

W15

Testing and CI for ML SystemsTest data, transformations and prediction contracts in addition to ordinary code.

10–12 hours per week0/5

Learn: Study unit, data, pipeline and model-behaviour tests.Practise: Create tests for schema drift, nulls, prediction shape and a quality floor.Build: Run lint, tests and a small smoke-training job in CI.Verify: Break one invariant and confirm CI blocks the change.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

An ML-aware CI pipeline with fast quality and contract gates.

W16

Monitoring, Drift and RollbackDetect when inputs or outcomes no longer resemble validated conditions.

10–12 hours per week0/5

Learn: Study service metrics, data drift, prediction drift, delayed labels and rollback triggers.Practise: Create reference and current batches with intentional drift.Build: Generate a monitoring report and define alert thresholds plus response owners.Verify: Simulate promoting and rolling back a bad model version.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

Version 4 with monitoring dashboard/report, alert policy and tested rollback.

MONTH 04 MILESTONEA containerised inference API with registry, CI checks, monitoring plan and rollback path.

Applied LLM Engineering and Evaluation

Add a language-model capability only where it improves the product and can be evaluated.

W17

Embeddings and Semantic RetrievalBuild retrieval from first principles before adding a generative layer.

10–12 hours per week0/5

Learn: Study chunking, embeddings, similarity, indexing and retrieval metrics.Practise: Create a small approved knowledge base and relevance-labelled queries.Build: Implement semantic search with citations to source passages.Verify: Measure recall at k and review missed or irrelevant results.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

A tested semantic retriever with labelled queries and citation metadata.

W18

RAG and Prompt BoundariesGround generation in retrieved evidence and make abstention possible.

10–12 hours per week0/5

Learn: Study retrieval-augmented generation, context assembly, prompts and structured outputs.Practise: Write adversarial and unanswerable support questions.Build: Generate draft answers from retrieved context with citations and refusal rules.Verify: Check every citation resolves to a supporting source passage.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

A RAG prototype that cites sources and abstains when evidence is insufficient.

W19

LLM Evaluation and GuardrailsReplace impressive demos with a repeatable quality gate.

10–12 hours per week0/5

Learn: Study task-specific test sets, groundedness, relevance, safety and latency/cost metrics.Practise: Label expected behaviour for normal, edge and adversarial cases.Build: Build an automated evaluation runner and human-review sheet.Verify: Compare two retrieval or prompt variants and document trade-offs.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

An evaluation report with dataset, rubric, baseline, improvement and known limits.

W20

Integrate the Assistant into the ProductShip the LLM feature with observability and a safe fallback workflow.

10–12 hours per week0/5

Learn: Study request tracing, feedback capture, rate limits and fallback UX.Practise: Test model failure, empty retrieval and slow response paths.Build: Add the assistant to the inference service with traces and user feedback.Verify: Demonstrate rollback or disablement without breaking core predictions.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

Version 5 with monitored RAG, citations, feedback and independent feature rollback.

MONTH 05 MILESTONEAn evidence-based support assistant with retrieval, citations, safety boundaries and repeatable evaluations.

Capstone and AI/ML Interview Readiness

Turn the work into credible evidence of modelling judgement and production ownership.

W21

Model Card, Ethics and Responsible UseDocument intended use, limitations, data risks and human oversight.

10–12 hours per week0/5

Learn: Study model cards, bias evaluation, privacy and failure impact.Practise: Evaluate performance across meaningful groups or slices.Build: Write a model card and define human-review triggers.Verify: Ask a reviewer to challenge one harmful or misleading scenario.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

A model card with intended use, evaluation, limitations and monitoring responsibilities.

W22

ML System DesignDesign the full path from data collection to feedback and retraining.

10–12 hours per week0/5

Learn: Study batch versus online inference, feature consistency, registries and retraining triggers.Practise: Design recommendation, fraud and ticket-routing systems under constraints.Build: Create one end-to-end architecture for the capstone.Verify: Defend freshness, latency, cost, quality and rollback choices.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

An ML system-design document with data, training, serving, monitoring and feedback loops.

W23

ML Coding and Case InterviewsPractise data manipulation, modelling and diagnosis under time pressure.

10–12 hours per week0/5

Learn: Review Python, SQL, probability, metrics and experiment interpretation.Practise: Complete timed data-cleaning, modelling and error-analysis exercises.Build: Prepare concise explanations of five capstone decisions and failures.Verify: Run one mock that changes the business metric midway.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

Three mocks covering Python/data, modelling/evaluation and ML system design.

W24

Final Capstone and PortfolioPresent measurable learning and production evidence rather than only notebook accuracy.

10–12 hours per week0/5

Learn: Review problem, data, baseline, experiments, serving, monitoring and LLM evaluation.Practise: Ask users or reviewers to test predictions and assistant answers.Build: Polish code, deployment, experiment history, dashboards and case study.Verify: Map each project artifact to skills requested in target roles.Document the decisions, evidence and remaining limitations for this week.

DELIVERABLE

A live AI/ML product, repository, experiment report, model card, architecture and five-minute demo.

MONTH 06 MILESTONEA deployed, monitored AI product with model card, case study and practical interview evidence.