roadmap
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08
Prepares you for AI and data science interviews, from coding rounds and statistics to explaining models end to end. Spans Python and SQL, ML and deep learning, LLM methods, and MLOps.
10279 of 10279 questions in your plan: 0 known, 0 in progress, 0 due, 0 skipped, 10279 new. 0 cut, outside your plan.
A data science loop puts its longest round somewhere a software loop does not: on the model walkthrough, a project you ran or a case the interviewer sets, taken from the business question through data, features, model and metric to what happened after launch. Around it sit a Python or SQL screen, a statistics conversation, often a take-home, and a behavioral round. Middle and senior loops usually add a design conversation about a recommendation, forecasting or retrieval system. The statistics round separates candidates most often, and its questions are precise. How many users does an experiment need to detect a two percent lift, and what changes when the metric is a ratio of two counts? Why does a launch that changed nothing still turn up one significant metric more often than not, once twenty are tested? Why does accuracy flatter a fraud model whose positive class is one case in a thousand? The modeling questions circle leakage: a scaler fit before the split, a target encoding computed on the whole set, a feature that only exists after the event it predicts. The AI half of the title gets the same scrutiny: how to score a retrieval system when no single answer is right, and when fine-tuning earns its cost over a better prompt. In the walkthrough, interviewers listen for the decision behind each step. Naming gradient boosting is where a junior answer ends; a senior one says why it suited this table, how the validation split mirrored the way the model would be used, and what the offline gain became after launch. The follow-up few candidates rehearse is the mismatch: the offline metric improved, the online one did not, and what you did next.
The plan opens on theory, and its two opening modules carry unequal weight. [Computer science fundamentals](/topics/found-cs-fundamentals) is here for the sections a data scientist actually meets, information theory behind cross-entropy most of all. [Math, statistics and experimentation](/topics/found-math-statistics) is the longest-lived module in the plan: linear algebra and optimization feed the model modules later, probability and inference feed the statistics round directly, and [experiment design](/topics/found-math-statistics-ab-testing) with causal inference is what the launch question in a walkthrough really asks. [Algorithms](/topics/found-dsa) follow for the coding screen, which most data science loops keep in a shortened form. Languages come next, [Python](/topics/lang-python) first and [SQL](/topics/lang-sql) because the data screen is usually a SQL screen. [Git](/topics/dev-git) and the testing disciplines sit right after them on purpose: the gap between a notebook and a codebase is where analysts lose a software-minded interviewer, and that gap is version control and tests. From there the plan follows the data as it grows. Its module on data that fits in one process comes in two layers: [the operations themselves](/topics/data-libraries-concepts), described without tying them to any one library, then [pandas](/topics/data-pandas), the API a take-home is usually written against, where a follow-up can turn on why an assignment through a chained selection left the original frame unchanged. The neutral layer turns, near its end, to what to do when the machine is too small, and the answer comes in two steps: the stores, from relational concepts to the [columnar warehouse](/topics/db-analytical-concepts) an analytics team lives in, then the [processing engines](/topics/data-processing-concepts) that split a job across workers, with the shuffles, skew and spills that come with it. Only now does the plan reach models, because a model is an argument about a table and the table had to come first. [ML fundamentals](/topics/found-ml-fundamentals) precede [deep learning](/topics/found-deep-learning), and the frameworks that implement both come after the ideas. The model hubs stretch turns pretrained models into something you can use, from fetching a checkpoint to [fine-tuning with the Trainer](/topics/ai-huggingface-finetuning), and ends on [Sentence Transformers](/topics/ai-sentence-transformers) for embeddings and rerankers, the bridge into the LLM half. That half runs from method to tool: [LLM methods](/topics/found-ai-llm-methods) first, where prompting, retrieval, agents, safety and the [scoring of output that varies between runs](/topics/found-ai-engineering-evaluation) are taught as ideas, along with the [vector indexes](/topics/found-embeddings-vector-indexes) that search the embeddings the previous stretch produced and the choice between [calling a hosted model and running open weights](/topics/found-ai-engineering-model-sourcing-hosted-vs-open). The lifecycle module follows, and it is about the work around a model rather than the model itself: [tracking an experiment](/topics/ai-mlflow-experiment-tracking) so it can be reproduced, [explaining a prediction](/topics/ai-shap) through SHAP values, and the cloud provider concepts the training runs on. [System design](/topics/found-system-design), with [ML system design](/topics/found-ml-system-design) inside it, supplies the senior round, and the plan closes on AI coding assistants, the [behavioral round](/topics/found-behavioral) and the shape of the loop itself.
Two weeks out, the plan is too long to read and the controls exist to shorten it. Set importance to must know alone and level to the band the job is posted at, leaving study order where the page opens it, by frequency; what survives is mostly statistics, ML fundamentals and SQL. Three months out, switch study order to by topics and read downward, keeping one end-to-end project growing beside the reading, because a walkthrough of a model you shipped survives follow-ups that a rehearsed one does not. Your background decides which half of the plan is the hard one. From statistics, economics or an analyst role, the math is already yours; spend the time on Python past the notebook, Git, tests and the deep learning and LLM stretch. From software engineering, the coding screen and the tooling are easy; spend it on pandas idioms, inference, experiment design and bias-variance, and on practising the walkthrough in words rather than code. Senior candidates should weight causal inference, the lifecycle module and ML system design; in the design round they will be asked for a serving latency, a retraining cadence and a plan for drift. The plan trains, tracks and explains a model; it does not ship one. Containers, serving infrastructure and the pipelines that retrain and redeploy on a schedule are the [MLOps Engineer roadmap](/roadmaps/mlops). If the job is building and scheduling the pipelines that deliver the data in the first place, the [Data Engineer roadmap](/roadmaps/data-engineer) is the one to open next.
A data science loop puts its longest round somewhere a software loop does not: on the model walkthrough, a project you ran or a case the interviewer sets, taken from the business question through data, features, model and metric to what happened after launch. Around it sit a Python or SQL screen, a statistics conversation, often a take-home, and a behavioral round. Middle and senior loops usually add a design conversation about a recommendation, forecasting or retrieval system.
The statistics round separates candidates most often, and its questions are precise. How many users does an experiment need to detect a two percent lift, and what changes when the metric is a ratio of two counts? Why does a launch that changed nothing still turn up one significant metric more often than not, once twenty are tested? Why does accuracy flatter a fraud model whose positive class is one case in a thousand? The modeling questions circle leakage: a scaler fit before the split, a target encoding computed on the whole set, a feature that only exists after the event it predicts. The AI half of the title gets the same scrutiny: how to score a retrieval system when no single answer is right, and when fine-tuning earns its cost over a better prompt.
In the walkthrough, interviewers listen for the decision behind each step. Naming gradient boosting is where a junior answer ends; a senior one says why it suited this table, how the validation split mirrored the way the model would be used, and what the offline gain became after launch. The follow-up few candidates rehearse is the mismatch: the offline metric improved, the online one did not, and what you did next.
The plan opens on theory, and its two opening modules carry unequal weight. Computer science fundamentals is here for the sections a data scientist actually meets, information theory behind cross-entropy most of all. Math, statistics and experimentation is the longest-lived module in the plan: linear algebra and optimization feed the model modules later, probability and inference feed the statistics round directly, and experiment design with causal inference is what the launch question in a walkthrough really asks. Algorithms follow for the coding screen, which most data science loops keep in a shortened form.
Languages come next, Python first and SQL because the data screen is usually a SQL screen. Git and the testing disciplines sit right after them on purpose: the gap between a notebook and a codebase is where analysts lose a software-minded interviewer, and that gap is version control and tests.
From there the plan follows the data as it grows. Its module on data that fits in one process comes in two layers: the operations themselves, described without tying them to any one library, then pandas, the API a take-home is usually written against, where a follow-up can turn on why an assignment through a chained selection left the original frame unchanged. The neutral layer turns, near its end, to what to do when the machine is too small, and the answer comes in two steps: the stores, from relational concepts to the columnar warehouse an analytics team lives in, then the processing engines that split a job across workers, with the shuffles, skew and spills that come with it.
Only now does the plan reach models, because a model is an argument about a table and the table had to come first. ML fundamentals precede deep learning, and the frameworks that implement both come after the ideas. The model hubs stretch turns pretrained models into something you can use, from fetching a checkpoint to fine-tuning with the Trainer, and ends on Sentence Transformers for embeddings and rerankers, the bridge into the LLM half.
That half runs from method to tool: LLM methods first, where prompting, retrieval, agents, safety and the scoring of output that varies between runs are taught as ideas, along with the vector indexes that search the embeddings the previous stretch produced and the choice between calling a hosted model and running open weights. The lifecycle module follows, and it is about the work around a model rather than the model itself: tracking an experiment so it can be reproduced, explaining a prediction through SHAP values, and the cloud provider concepts the training runs on. System design, with ML system design inside it, supplies the senior round, and the plan closes on AI coding assistants, the behavioral round and the shape of the loop itself.
Two weeks out, the plan is too long to read and the controls exist to shorten it. Set importance to must know alone and level to the band the job is posted at, leaving study order where the page opens it, by frequency; what survives is mostly statistics, ML fundamentals and SQL. Three months out, switch study order to by topics and read downward, keeping one end-to-end project growing beside the reading, because a walkthrough of a model you shipped survives follow-ups that a rehearsed one does not.
Your background decides which half of the plan is the hard one. From statistics, economics or an analyst role, the math is already yours; spend the time on Python past the notebook, Git, tests and the deep learning and LLM stretch. From software engineering, the coding screen and the tooling are easy; spend it on pandas idioms, inference, experiment design and bias-variance, and on practising the walkthrough in words rather than code. Senior candidates should weight causal inference, the lifecycle module and ML system design; in the design round they will be asked for a serving latency, a retraining cadence and a plan for drift.
The plan trains, tracks and explains a model; it does not ship one. Containers, serving infrastructure and the pipelines that retrain and redeploy on a schedule are the MLOps Engineer roadmap. If the job is building and scheduling the pipelines that deliver the data in the first place, the Data Engineer roadmap is the one to open next.