Get Up to 40% OFF New-Season StylesMenWomen * Limited time only.

10 Real-World Generative AI Projects Every Data Science Student Should Build

10 Real-World Generative AI Projects Every Data Science Student Should Build

The vast majority of data science students can train a classifier, clean a dataset and build a dashboard. Very few can actually go on to design a generative AI system that can solve a business problem in a robust manner, cope with incomplete data, protect the data and be able to explain what it’s done. That’s a huge gap and it’s absolutely true that employers are increasingly looking for people who can take the LLMs, the retrieval systems and the NLP and turn them into real applications rather than just give them a prompt and say “there you go”. 

Generative AI projects are only going to be of any use if they’re actually solving a workflow problem. That might be answering questions against a corpus of trusted documents, extracting structured data from unstructured text, classifying customer feedback, generating some kind of content or assisting with a human decision. A good AI project in a portfolio will have some real data (or at least some fairly realistic data), a clear objective, a way of measuring success, a UI (or API) that you can actually use and a very honest discussion of the failure modes. 

Here’s a list of 10 practical generative AI projects for data science students (for the aspiring AI engineer and the self-taught). Each one is designed to showcase the skills that a recruiter or hiring manager would be looking for: Python, SQL, data wrangling, NLP, LLM application design, RAG, evaluation, deployment and responsible AI. You’ll also learn how to pick a project, what architecture to use, what metrics to care about and how to actually present your work. 

Current landscape: from models to systems 

Generative AI has long moved past the era of the basic chatbot. It’s now being used in combination with search, databases, business apps and human review workflows to automate huge volumes of knowledge work. The real question isn’t any longer ‘which model should I use?’ It’s ‘how do I build a system that will produce useful, safe and verifiable outputs in a given business context?’ 

For a data science student that’s completely changed the bar on what constitutes a good project. A notebook that calls an LLM API and returns an answer is a demo. A project that takes in documents, finds the relevant evidence, generates a response, validates the output, logs the results and then exposes the system via a simple interface is the kind of project that would actually be usable. 

Modern AI portfolios should be showing off the fundamentals of ML and product thinking. You need to be able to articulate a user problem, get the data into the right state, pick the right model or API, assess the quality of the output, deal with errors and have a good understanding of the limits of the system. Generative AI in absolutely no way gets rid of the need for data science. It actually underlines even more the importance of good data handling, good measurement and good domain understanding. 

Major changes: why GenAI projects are different 

Traditional ML projects are of the form ‘map an input to a prediction’ (e.g. predict churn from historical data). Generative AI projects are a completely different beast in that they are producing new content (answers, summaries, code, reports, classifications, structured data etc.) in response to a natural language instruction and context. That in itself creates a very different set of engineering requirements. 

  • Context is everything. The answer the LLM produces is entirely a function of the instructions, the documents it has retrieved, the conversation history and the output constraints. 
  • Evaluation is probabilistic. Two valid answers could easily be worded differently so exact match accuracy on its own is rarely a good enough metric. 
  • Grounding is absolutely vital. For business use in particular it’s essential that the responses can point to and trace back to the correct source rather than just drawing on model memory. 
  • Interfaces matter. A useful project will have an API, a web interface, a batch pipeline etc. It doesn’t have to be a Jupyter notebook. 
  • Governance is the absolutely key thing. You have to be thinking about privacy, bias, security, IP and harmful content from day one. 

Retrieval Augmented Generation (RAG) is very much the thing in this context as it allows the LLM to be connected to external knowledge. Rather than the model having to go through its training data to find the information it needs, the application will retrieve the relevant source passages and then ask the model to generate its answer on the basis of those passages. That obviously can lead to a more relevant answer and better traceability, but in no way removes the potential for error. You can still end up with the wrong answer as a result of poor retrieval, conflicting sources or a bad prompt. 

Hugging Face Transformers offer both pipelines and the generate() API for text generation models. This covers the full range of text generation tasks from the very simple to classification, summarisation, translation and question answering. The pipelines and the generate() API are extremely useful for students who want to play with open models either on their own machine or in the cloud. API based models can be a good option if all you’re really after is rapid prototyping and having a managed infrastructure rather than actually owning the model. [web:11][web:14][web:15] 

Practical impact on professionals 

Generative AI is going to completely transform the way data professionals do their day to day jobs. Rather than spending all their time manually going through documents, writing the same old reports, labelling text, searching across systems etc they can be designing AI assisted workflows which will accelerate all of those tasks. The human is still ultimately responsible for the high impact decisions, for validating the outputs and for maintaining the system.

A support analyst could have spent their career going through hundreds of tickets and worked out what the common problems were. A good GenAI project would be able to categorise the tickets, extract the product names and the issue types, identify the themes and then route the most urgent ones to the appropriate team. The analyst would then look at the exceptions and make the operational decisions on the basis of the structured output.

This is the graduate that the AI delivery lifecycle rewards – the one who can go right through the whole thing from problem definition, data access, model selection, prompt/retrieval design, evaluation, deployment, monitoring and then of course the communication with all the various stakeholders. The graduate who can not only explain how the model works but can also explain when it’s not appropriate to use it is the one that will really stand out.

Business implications

The business value of generative AI is in reducing effort, improving consistency, increasing speed and/or enabling better decisions. Value is only realised when the output is good enough for the task and where humans can step in where an error would be a problem.

A Q&A assistant on a document would save a huge amount of time that would otherwise be spent trawling through policy docs. A customer feedback classifier would be a massive time saver for a product team in terms of spotting issues as they emerge. Report generation would be a huge time saver for analysts (of course), though you’d be mad to have it automatically publishing any financial or compliance conclusion without some form of human review.

When you’re looking at a project work out the difference between a high volume low risk job and a high risk high stakes decision. Summarising public product reviews is going to be a lot lower risk than coming up with medical advice, legal advice, a hiring decision or a credit decision. The more serious the output the more you need human approval, audit trails and clear disclaimers.

Technical considerations

Every good generative AI project should start with a very specific use case and a very clear definition of what success looks like. You shouldn’t be out there trying to build a ‘general AI assistant’. You need to define the user, the input, the output, the constraints and how you’re going to measure success before you write a single line of code.

  • Define the task: e.g. “Classify the customer reviews into product issues, delivery issues, pricing issues and praise”, not “Analyse the customer feedback”.
  • Prepare the data (redact/mask any personally identifiable information as appropriate, document the source, create a test set that is representative of real world use).
  • Choose the right architecture. For simple tasks use prompt engineering. For questions that require a lot of knowledge use RAG. Only fine tune if the prompt based approaches aren’t good enough in terms of quality or consistency. Use classical ML if all you really need is a deterministic classifier.
  • Design for evaluation: measure the task specific metrics (retrieval precision, answer faithfulness, classification F1-score, latency, cost and human review agreement).
  • Build the guardrails: validate the inputs, constrain the outputs to the correct formats, log the requests and responses, deal with errors appropriately and provide a human review pathway for any sensitive cases.

For local experimentation you can have Hugging Face Transformers load a pre-trained model via its pipeline interface (which brings the pre-processing, model inference and post-processing all in one). The text generation pipelines take a prompt or a chat-style input and you can control things such as max new tokens and sampling. [web:11][web:12][web:14] Structured output and function calling are useful if you want the LLM to return some structured fields rather than freeform text. Obviously you’ll want to validate the returned JSON against a schema before you store it or act on it.

Risks, limitations, and governance

Generative AI brings with it a whole host of risks that traditional software projects may not have had to contend with. It can generate information that is extremely confident and yet entirely wrong. It can leak data if it’s connected to the wrong data sources. It can propagate bias and it can be fooled by adversarial prompts. NIST have a generative AI risk profile which covers confabulation, privacy, information integrity, security, intellectual property, bias and integration risks. [web:2][web:3]

NIST’s AI Risk Management Framework is built on four functions: Govern, Map, Measure and Manage. It’s voluntary but provides a very useful structure for student projects in that you essentially document the intended use, identify the users and data that are affected, measure the performance and harms and then define what you do in the event of a failure. [web:1][web:4][web:5]

  • Hallucination: Never present any generated content as a fact unless you have source evidence or have actually looked at it.
  • Privacy: don’t upload any confidential, personal or proprietary data to a public model or service without the appropriate levels of permission and protection in place.
  • Prompt injection: Treat the retrieved documents, the user input and any third party content as untrusted; separate the instructions from the data and validate the requested action.
  • Bias and fairness: test performance across the relevant user groups, languages, product categories and/or document types as appropriate.
  • Overreliance: design the system to support a human decision rather than to remove the accountability for decisions which have significant consequences.

Cost and latency: track token usage, response time and infrastructure cost (this is especially important if you’re looking at a project for production).

10 real-world generative AI projects

The projects listed below are in order from most accessible to most advanced. Each of them could form the basis of a capstone project, a portfolio case study or a piece of interview material if you include the data prep, the evaluation, the deployment and the limitations.

1. AI-powered customer support knowledge assistant

Build a RAG based assistant which will answer customer support queries by pulling in the relevant bits from a curated set of FAQs, product manuals, return policies and troubleshooting guides. It will retrieve the most relevant passages, construct a succinct answer and provide links to the source documents.

Use a public FAQ dataset or create a realistic synthetic knowledge base for a made up e-commerce company. Do document chunking, embeddings, vector search, prompt templates and citation display. The passages returned must contain the answer and the generated response must be true to those passages.

  • Skills demonstrated: RAG, embeddings, vector databases, prompt engineering, evaluation, Streamlit or FastAPI.
  • Business value: reduces the number of repetitive support queries and enables agents to find the approved answers a lot faster.
  • Important limitation: The assistant must be absolutely clear when it cannot find a reliable answer and must never make up a policy.

2. Resume-to-job-description matching assistant

Create a system that takes a CV and compares it to a job description to produce a structured fit analysis. The analysis should include the matching skills, the skills that are missing, the relevant experience and some suggested improvements. Use function calling / structured output to return a JSON and then display the result in a simple web interface.

This isn’t just some simple text generation. It’s about parsing PDFs and DOCXS, extracting entities from them, coping with the sheer variety of resume formats out there and the fact that people will always invent experience. You need a confidence score and it has to be crystal clear that this is decision support and not an automated hiring decision.

  • Skills demonstrated: Document parsing, NLP, structured outputs, prompt constraints, responsible AI.
  • Business value: Helps the recruiter or career coach to cut through the pile of applications.
  • Important limitation: Don’t use the system in any way at all for actual hiring, rejection or ranking decisions without some form of human review.

3. Customer-review insight and theme classifier

Build an NLP and generative AI pipeline that takes in product reviews, classifies them into themes, extracts the product features and sentiment and then generates a weekly insight report. It could for instance identify the various issues that keep reappearing around delivery times, battery life, sizing and packaging.

Use a hybrid approach: a traditional classifier or embedding based clustering will deal with the repeatable categories and the LLM will be used to summarise the new themes and produce an executive summary. Benchmark the LLM only approach against the hybrid approach using labelled examples, precision, recall, F1 and a qualitative review.

  • Skills demonstrated: Text classification, topic modelling, sentiment analysis, data visualisation, report generation.
  • Business value: takes the unstructured feedback and turns it into real product and operations insight.
  • Important caveat: the generated summaries have a well-documented tendency to overemphasise the odd review. Always remember that the themes you’re seeing may not be reflective of the actual sample sizes or the source material.

4. Meeting-notes summarizer and action-item extractor

Create an app that takes a meeting transcript and produces a properly structured summary with decisions, action items, owners, deadlines, risks and open questions. Take either a transcript from the public meeting corpus or a synthetic transcript and return both a readable summary and a machine readable JSON.

The major issue is faithfulness. The system has to be able to tell the difference between what was actually said and what is in fact an assumption or a recommendation. Add a validation layer to pick out the action items which have neither an owner nor a deadline and allow the user to edit the final output prior to saving it.

  • Skills demonstrated: Speech to text integration, summarisation, information extraction, structured output, workflow automation.
  • Business value: reduces the need for manual note taking and improves follow up on decisions.
  • Important limitation: Don’t store any confidential meeting recordings or transcripts without consent and with appropriate access controls.

5. SQL query assistant with schema-aware guardrails

Build an assistant that takes a natural language business question and translates that into SQL for a given database schema. So if a user asked “Show me monthly revenue by product category for the last 6 months” the system would generate the appropriate SQL query against the sales database.

Use a small, well-documented sample database with very clear table relationships. Include the schema in the prompt, validate the generated SQL, restrict writes (so you’re only doing SELECTS), run it in a sandbox, show the results of the query and then give an explanation of those results. That’s a perfect example of the connection between generative AI, SQL and analytics.

  • Skills demonstrated: SQL, schema design, prompt engineering, validation, data visualisation and API development.
  • Business value: Makes data exploration accessible to the non-technical business user.
  • Important limitation: you must never allow unimpeded database access. The SQL that is produced must be validated and access must be restricted.

6. Intelligent document-processing pipeline

Design a pipeline to extract the structured data from the unstructured documents (invoices, purchase orders, application forms, insurance claims etc). The approach would be to convert the document to text, identify the fields of interest, use an LLM to normalise the values and then validate the output against the business rules.

For instance you’d pull out the invoice number, vendor name, invoice date, line items, currency, total amount and payment due date. You’d have rule based checks to catch invalid dates, totals that don’t match, missing required fields etc. That project really does underline the point that generative AI is at its most powerful when it’s combined with some form of deterministic validation.

  • Skills demonstrated: OCR / document parsing, information extraction, data validation, ETL, exception handling.
  • Business value: reduces the need for manual data entry and ensures greater consistency of documents in document intensive processes.
  • Important caveat: all the financial and legal docs will always require human approval (in particular where the values that are extracted are in any way going to result in a payment or a compliance decision).

7. Personalised learning-path generator

Build a recommendation and content generation system which will create a personalised learning plan for the learner based on their current skills, their desired role, how much time they have available to study and their preferred learning style. The system will be able to map skill gaps to modules, produce a weekly plan and generate practice questions from the approved course content.

Use a structured skills taxonomy rather than the model going off and inventing courses and content from scratch. Take the approved learning content, generate the plan in a very specific format and then let the learner go and play with their own constraints (how many hours a week, what job role they’re targeting etc). That’s very much a project for the education space, HR tech and upskilling.

  • Skills demonstrated: Recommendation logic, RAG, personalisation, structured generation, user interface design.
  • Business value: allows the learner to progress from a broad career aspiration to a very specific, time-bound study plan.

Important caveat: the plan must be in line with the learner’s choice. It must not make any false or misleading claims as to employment outcomes.

8. Multilingual product-content generator with brand controls

Create a controlled content generation system to produce product descriptions, social media captions and email drafts in a number of languages. It should use the approved product attributes, tone of voice rules and prohibited claims and go through a review process prior to publication.

A robust version includes a template driven prompt, a glossary of approved terminology, language specific validation and a human approval queue. Facts are checked against the product catalogue, readability is checked, brand rule breaches are checked for and the editor revision rate is monitored.

  • Skills demonstrated: Prompt engineering, multilingual NLP, content moderation, workflow design, quality evaluation.
  • Business value: gets the content out the door a hell of a lot faster, whilst keeping the brand and the facts spot on.
  • Important limitation: Do NOT allow the model to make any unsupported product, health, financial or legal claims. claims.

9. AI-powered incident-triage assistant

Build a system that takes incident reports, support tickets or system logs and works out the severity, likely impacted service, probable root cause category and recommended next action. It should pull in the relevant runbooks and past resolved incidents prior to recommending any action.

Use a realistic synthetic dataset (real operational logs would contain sensitive information). Include a confidence score, source citations and an escalation rule for high severity incidents. The most useful thing you can produce is a concise, evidence-based starting point for an on-call engineer. The fix itself isn’t the point.

  • Skills demonstrated: Log analysis, classification, RAG, retrieval evaluation, operational AI, alerting.
  • Business value: Helps support and engineering teams to prioritise issues and to get to the relevant historical context a lot quicker.
  • Important limitation: any automated actions on production systems must be subject to strict approval controls and audit logging.

10. Responsible AI audit and evaluation dashboard

Build an evaluation dashboard for one of your own generative AI apps. Have it collect: test prompts, expected answers (or reference outputs), model responses, retrieved sources, latency, token usage, cost and human ratings. Then visualise quality across the various dimensions of quality: category, language, question difficulty and failure type.

This project makes evaluation a first class citizen of the AI system. You get the standard metrics of answer relevance, factuality (against the retrieved evidence), citation accuracy, refusal quality, harmful output rate and cost per successful request. A dashboard that shows you where it’s failing is almost always far more persuasive than one that just shows you a successful demo.

  • Skills demonstrated: Evaluation design, data collection, analytics, visualisation, quality assurance, responsible AI.
  • Business value: Gives the teams the evidence they need to work out whether an AI feature is ready to go wider.

Important caveat: automated metrics are useful but incomplete. For any high impact use case you’re going to need some level of structured human review.

How to choose the right project

Pick one project that’s appropriate for the role you’re targeting and that you can do to a high standard. A focused project with good documentation, tests, evaluation results and a deployed demo is vastly more valuable than five notebooks that you’ve never actually completed.

  • For data analyst roles: the SQL assistant, the customer review insight project or the evaluation dashboard.
  • For data science roles: focus on review classification, document processing, incident triage or evaluation dashboards.
  • For AI engineering roles: RAG assistants, structured output apps, document pipelines and deployed APIs.
  • For product roles: the learning path generators, the content generators, the support assistants and the workflow tools (all of which have some level of human review).

Skills professionals should develop.

A job-ready generative AI portfolio requires more than prompt-writing. Build competence across the following areas:

  • Python and the basics of software development (modular code, version control, testing, logging and environment management).
  • Data engineering: data cleaning, schema design, SQL, APIs, document parsing and privacy aware data handling.
  • NLP and machine learning: embeddings, classification, clustering, evaluation and classical baselines.
  • LLM application development: prompt design, RAG, function calling, structured outputs, caching and model selection.
  • Evaluation and observability: test sets, human review, quality metrics, latency, cost and failure analysis.
  • Deployment: Streamlit or Gradio for demos, FastAPI for services, Docker for reproducibility and cloud platforms for hosting.

Responsible AI: access controls, privacy, bias testing, content safeguards, audit trails and human oversight.

What this means for TrainingYA learners

For anyone looking to build a career in data and AI the best way to do it is to have a good solid foundation in analytics skills and then go on to do some applied generative AI. The core analytics skills (strong SQL, Python, Excel, Power BI, data analysis etc.) are still absolutely fundamental as GenAI systems are ultimately built on data, well defined business questions and measurable outcomes.

TrainingYA’s industry focused approach to learning (practical projects, mentoring, career support) is very much in keeping with this portfolio first approach. The learner is able to build a SQL query assistant, analyse the customer feedback, visualise the results in Power BI and then be able to articulate the limitations of the system they’ve built. That’s applied capability rather than just theoretical knowledge.

Start with one end to end project, document the business problem and the outcome, publish the code and a short case study and then iterate on the basis of the evaluation. That discipline alone will go a long way to making your generative AI projects credible to recruiters and actually useful in an organisation.

The best projects are ones that actually solve a very specific, tangible problem and demonstrate end to end capability. A RAG based support assistant, a customer review insight classifier, a SQL query assistant, a document extraction pipeline, a meeting notes summariser and an AI evaluation dashboard would all be good examples of that.

No. The vast majority of good student projects will be using a pretrained model via an API or a pretrained open source model. You should be starting with data prep, retrieval, prompt design, evaluation, deployment and responsible use. Fine tuning is only ever appropriate when you can’t achieve the required quality using a prompt based approach.

RAG is retrieval augmented generation. It works by retrieving relevant information from an approved knowledge source and then having the LLM generate an answer on the basis of that. It’s most commonly used for document Q&A, internal knowledge assistants and support bots.

Python is by far the most popular language for generative AI projects. You need to know SQL, Git, pandas, FastAPI or Streamlit (whichever you prefer), vector databases, Hugging Face Transformers and at least one cloud platform. Basic Docker is a useful thing to have as well.

Use a mix of automated and human evaluation. Depending on the task you’re looking at, you’ll want to measure retrieval relevance, answer faithfulness, classification precision and recall, F1-score, latency, cost, citation accuracy and human review agreement. Always test the failure cases as well as the ideal prompts.

Only with proper authorisation. Don’t upload any confidential, personal or proprietary data to any public service without appropriate safeguards in place. For a portfolio project, you’re probably best off using public data, synthetic data or anonymised sample data.

Yes. A properly scoped project such as a document question-answering assistant, a customer feedback analytics system or a SQL assistant will quite clearly satisfy capstone requirements and at the same time demonstrate real AI engineering skills.

The most common error is to build a prompt demo without any evaluation or actual deployment. The recruiter wants to see how you’ve dealt with data quality, errors, privacy, testing, UX and limitations.

They don’t just want to see that you’ve called an LLM API.

Publish a GitHub repo with full documentation, an architecture diagram, evaluation results, a simple demo and a short case study. The case study should detail the business problem, the approach taken, the metrics used, the limitations and what could be done differently in the future.

Build a task-specific application. A specialist assistant which can answer questions from a set of approved documents, extract invoice fields or classify support tickets is far easier to assess and far more useful than a general-purpose chatbot with no real business purpose.

Conclusion

Generative AI projects are only going to be of real value if they show the complete journey from a genuine business problem through to a robust, evaluated and responsibly designed application. For data science students, the best portfolio pieces are those that integrate LLMs with data engineering, retrieval, structured outputs, evaluation, deployment and human-in-the-loop.

Start with one project that’s relevant to your career path: a RAG knowledge assistant for AI engineering, a customer review analytics pipeline for data science or a SQL query assistant for an analytics role. Build it from the ground up, assess its performance in a real-world scenario (don’t just do a paper exercise), document its shortcomings and be absolutely clear that a human is still in control of the decisions that matter.

The people who will benefit most from generative AI aren’t just the prompt users. They’re the dataaware builders who can take a generative AI project and turn it into a real system that makes better decisions, saves time and is trustworthy in the real world.

Share this post

Leave a Reply

Your email address will not be published. Required fields are marked *


Chat with Counsellor