The best AI and ML projects for 2026 pair a real dataset with an honest baseline and a clear metric: house prices, customer churn, spam, fraud, recommendations, image and text classification, forecasting, then a RAG question-answering app with an evaluation set. Ship at least one behind an API or demo.
Hiring managers screen ML portfolios for three things: a real dataset, a number that beats a simple baseline, and evidence that you can ship.
- 1QuestionWho decides what from this model?
- 2Data and licenceReal source, check the licence
- 3BaselineSimple model first, write the number
- 4ImproveFeatures, tuning, a better model
- 5Evaluate honestlyRight metric, no leakage, error analysis
- 6Ship and documentAPI or demo plus a README
The 10 projects
House price prediction
The question: Which features drive sale price, and how accurately can we predict it?
- Start with
- Kaggle: House Prices, Advanced Regression Techniques. For a zero-download start, scikit-learn ships the California housing set.
- Steps
- Explore missing values and skew; log-transform the target.
- Fit a linear baseline and record its error.
- Engineer features (age of house, total area, quality bins).
- Try gradient boosting with cross-validation.
- Plot feature importance and the five worst predictions.
- Measure
- RMSE on log price, compared with the linear baseline.
- Ship
- Notebook plus a one-page README with the error table.
- Why it counts
- Shows the full supervised-learning loop on messy tabular data.
Built a house-price model (RMSE [x] vs [y] baseline) using gradient boosting and cross-validation; documented the top drivers and failure cases.
Customer churn model served as an API
The question: Which customers are likely to leave, and what is a sensible threshold to trigger a retention offer?
- Start with
- Telco Customer Churn (Kaggle): about 7,000 customers with contract, tenure and service columns.
- Steps
- Split train and test before any preprocessing.
- Baseline with logistic regression; then a tree model.
- Handle imbalance with class weights and compare with resampling.
- Choose a threshold from a cost table (offer cost vs lost customer value).
- Wrap the model in a FastAPI
/predictendpoint and add one test.
- Measure
- ROC-AUC and precision-recall at your chosen threshold, not accuracy.
- Ship
- Repo with a running API, a request example and the threshold reasoning.
- Why it counts
- Joins modelling with deployment, the combination most junior portfolios lack.
Trained and deployed a churn classifier (ROC-AUC [x]) behind a FastAPI endpoint; set the decision threshold from retention-offer cost vs customer value.
SMS spam classifier
The question: Can a lightweight model separate spam from genuine messages, and where does it fail?
- Start with
- UCI SMS Spam Collection.
- Steps
- Clean text and inspect class balance.
- TF-IDF plus Naive Bayes or logistic regression baseline.
- Swap in sentence embeddings and compare.
- List the ten most confident mistakes and explain them.
- Measure
- Precision and recall for the spam class; report the confusion matrix.
- Ship
- Notebook with an error-analysis table.
- Why it counts
- A clean first NLP project and a baseline you reuse in #7 and #9.
Built an SMS spam classifier (precision [x], recall [y]) and compared TF-IDF with embedding features; analysed top misclassifications.
Credit-card fraud detection
The question: How do we catch rare fraud without flooding analysts with false alarms?
- Start with
- Credit Card Fraud Detection (Kaggle, ULB): 284,807 transactions, 492 of them fraud (0.172%).
- Steps
- Split by time or stratify so fraud appears in both sets.
- Baseline, then class weights, then resampling (SMOTE) on the training set only.
- Plot the precision-recall curve.
- Pick an operating point from the cost of a missed fraud vs a false alert.
- Measure
- PR-AUC and recall at fixed precision. Accuracy is meaningless here (99.8% by predicting never fraud).
- Ship
- Notebook plus a short decision memo on the operating point.
- Why it counts
- Demonstrates you understand imbalance and business trade-offs.
Built a fraud detector on 0.17% positive data; selected an operating point from precision-recall and alert-review cost.
Movie recommender with ranking metrics
The question: Which five movies should we show this user, and how do we know the list is good?
- Start with
- MovieLens (GroupLens): start with the 100K set, move to 1M once it works.
- Steps
- Build a popularity baseline.
- Item-item cosine similarity (collaborative filtering).
- Matrix factorisation (ALS or SVD).
- Hold out each user's latest ratings for testing.
- Describe the cold-start problem and one fix.
- Measure
- Precision@5 and NDCG@10 against the popularity baseline.
- Ship
- Notebook and a small demo that takes a user ID and returns five titles.
- Why it counts
- Introduces ranking metrics, which differ from classification metrics.
Built a recommender (NDCG@10 [x] vs popularity [y]) using matrix factorisation on MovieLens; evaluated with time-based holdout.
Image classifier with transfer learning
The question: Can a pretrained vision model be adapted to a small custom image set, and what does it get wrong?
- Start with
- Fashion-MNIST (70,000 images, 10 classes) as a baseline, then 50 to 100 photos per class you collect yourself.
- Steps
- Train a small CNN baseline on Fashion-MNIST.
- Fine-tune a pretrained backbone on your own small set.
- Add augmentation and compare.
- Inspect misclassified images and the confusion matrix.
- Deploy a drag-and-drop demo on Hugging Face Spaces.
- Measure
- Accuracy and per-class F1 on a held-out set you did not tune on.
- Ship
- Live demo link, model card and error gallery.
- Why it counts
- Shows you can adapt a pretrained model, not only train from scratch.
Fine-tuned a pretrained vision model on a custom [n]-image dataset (macro-F1 [x]); deployed a public demo and documented failure cases.
Review sentiment: TF-IDF vs a fine-tuned transformer
The question: Is the extra cost of a transformer worth it over a bag-of-words baseline for this task?
- Start with
- IMDB reviews on Hugging Face (50,000 labelled reviews).
- Steps
- TF-IDF plus logistic regression baseline.
- Fine-tune a small transformer (DistilBERT class) for one to three epochs.
- Compare accuracy, training time and inference cost.
- Collect 20 reviews both models get wrong (sarcasm, negation) and categorise them.
- Measure
- F1 and accuracy, plus a table of training time and latency per model.
- Ship
- Comparison table and an error-category chart.
- Why it counts
- Teaches the trade-off thinking behind real model choices.
Compared TF-IDF with a fine-tuned transformer on 50k reviews (F1 [x] vs [y]); quantified the cost and latency difference and categorised failures.
Document Q&A (RAG) with an evaluation set
The question: Does the assistant answer only from the documents, and can we measure when it is wrong?
- Start with
- Public PDFs you can legally use, for example government FAQs, a university handbook or open product manuals. Write 30 questions with reference answers yourself.
- Steps
- Chunk documents by headings, not fixed length alone; store the source page with each chunk.
- Embed and index; return the top-k with citations.
- Add hybrid search (BM25 plus dense, merged with reciprocal rank fusion) and a cross-encoder reranker; compare each step.
- Score with Ragas: faithfulness, answer relevancy, context precision, context recall.
- List five failures, fix one, and re-measure.
- Measure
- The four Ragas metrics plus your own pass or fail on the 30 questions, before and after each change.
- Ship
- Demo with visible sources, an eval table and a failure write-up.
- Why it counts
- Most RAG demos show no measurements. An eval table is what separates engineering from a demo.
Built a cited document Q&A system and a 30-question evaluation set; hybrid retrieval plus reranking raised context precision from [x] to [y] and faithfulness to [z].
LLM vs classical baseline on real complaints
The question: Should complaint routing use a prompted LLM, a fine-tuned small model or TF-IDF, once cost and latency count?
- Start with
- CFPB Consumer Complaint Database: public complaints with a text narrative and a product label.
- Steps
- Sample 5,000 complaints across six product categories.
- Baseline: TF-IDF plus logistic regression.
- Few-shot prompt an LLM with structured output; add a fine-tuned small model.
- Measure accuracy, cost per 1,000 items and median latency for each.
- Write a one-paragraph recommendation.
- Measure
- Macro-F1 per approach, plus cost and latency columns in one table.
- Ship
- Comparison table and recommendation memo.
- Why it counts
- Shows model-selection judgement, a core AI-engineer skill in 2026.
Benchmarked TF-IDF, a fine-tuned small model and a few-shot LLM on [n] complaints; recommended an approach using macro-F1, cost per 1,000 items and latency.
Weekly demand forecast with a naive baseline
The question: How many units of the top products will sell next month, and how much better is the model than a naive guess?
- Start with
- UCI Online Retail II: about 1.07 million UK retail transactions from 2009 to 2011, licensed CC BY 4.0.
- Steps
- Remove cancellations (invoice numbers starting with C) and aggregate to weekly units for the top 20 products.
- Baselines: last value and seasonal naive.
- Fit a statistical model and a gradient-boosted model with lag features.
- Use rolling-origin (time-based) validation, never a random split.
- Measure
- MAPE or sMAPE against the seasonal naive baseline on rolling windows.
- Ship
- Forecast chart per product and a table of error vs baseline.
- Why it counts
- Time-series validation is a frequent interview topic and is easy to get wrong.
Forecast weekly demand for 20 products using rolling-origin validation (sMAPE [x] vs seasonal-naive [y]) on 1M+ transactions.
14 more project ideas
Shorter briefs for when you want a different angle. Apply the same rules: a real source, a baseline, a number and a README.
Loan approval predictor with explainability
Beginnerscikit-learn, SHAP, Streamlit
DataKaggle loan prediction or UCI credit data
Stand-outPer-applicant explanation of why a loan was declined
Air-quality forecasting for one city
Intermediatepandas, Prophet or LightGBM
DataCPCB data via data.gov.in or Kaggle
Stand-outRolling-origin validation and a 3-day public forecast page
Resume-to-job matcher with embeddings
Intermediatesentence-transformers, FAISS
DataPublic resume and job-posting datasets
Stand-outPrecision@5 and a bias check across genders and colleges
Plant disease detector for Indian crops
IntermediatePyTorch, transfer learning, Streamlit Community Cloud
DataPlantVillage plus 50 photos you take
Stand-outAccuracy on your own photos vs the clean dataset
Speech-to-text notes for lectures
IntermediateWhisper (open source), Python
DataYour recorded lectures (with permission)
Stand-outWord error rate on a 10-minute sample you transcribe by hand
Hindi or Tamil sentiment classifier
IntermediateHugging Face Transformers, IndicBERT
DataIndicNLP or Kaggle Indian-language reviews
Stand-outCompares a multilingual model with a language-specific one
Image captioning demo
AdvancedBLIP or a small vision-language model, Gradio
DataCOCO sample or your photos
Stand-outHuman rating of 50 captions against the model's confidence
Anomaly detection in server or sensor logs
IntermediateIsolation Forest, autoencoder
DataNASA bearing dataset or synthetic logs
Stand-outPrecision and recall on planted anomalies
Music genre classifier
Beginnerlibrosa, CNN
DataGTZAN dataset
Stand-outConfusion matrix explaining which genres overlap and why
Handwritten digit app with your own drawings
BeginnerKeras, a canvas web demo
DataMNIST plus digits you draw
Stand-outAccuracy drop on real drawings and what fixed it
Customer lifetime value regression
Intermediatepandas, gradient boosting
DataUCI Online Retail II
Stand-outBusiness rule: who gets a retention offer, with expected ROI
Fine-tune a small LLM for a narrow task
AdvancedHugging Face PEFT or LoRA, Colab
DataYour own 500-example dataset
Stand-outBefore and after eval set with a classical baseline
Semantic search over course notes
Intermediateembeddings, FAISS, Streamlit
DataYour own notes and PDFs
Stand-outRecall@5 on 30 queries vs keyword search
Object detection on a traffic clip
AdvancedYOLO (open weights), OpenCV
DataA short video you film legally
Stand-outCounts per minute and a note on false positives
README template that gets read
- First line: the question and the answer in one sentence.
- Data source and licence.
- Method in five bullets, baseline first.
- One chart and one headline number.
- Limits: what the data or model cannot tell you.
- Run it in one command (
docker compose upormake run).
Mistakes that make a portfolio look junior
- Only toy datasets (Iris, Titanic, MNIST) with no deployment or twist.
- Reporting accuracy on imbalanced data.
- Leakage: scaling or resampling before the train-test split, or random splits on time data.
- No baseline, so nobody can tell whether the model helps.
- A notebook with no run instructions and no stated limits.
Tools used in industry for these projects
| Job | Tool |
|---|---|
| Data and modelling | pandas, scikit-learn, XGBoost, LightGBM, PyTorch |
| Serving | FastAPI, Docker, Hugging Face Spaces, Streamlit |
| Retrieval | FAISS, Chroma, Qdrant, Elasticsearch BM25 |
| LLM evaluation | Ragas, custom golden sets, trace viewers |
| Experiment tracking | MLflow, Weights & Biases |
Frequently asked questions
What are good machine learning projects for beginners?
Predicting house prices, customer churn and spam messages are strong starts because they use real tabular or text data, have clear metrics and take under a day each.
How many ML projects do I need for a job application?
Two or three finished and documented projects are enough. One should be deployed behind an API or demo, and one should show an honest evaluation.
Do beginner ML projects need deep learning?
No. Most hiring screens test tabular modelling, evaluation and deployment. Add one deep learning or LLM project once the basics are solid.
Which AI project impresses most in 2026?
A retrieval-augmented question-answering app that includes a measured evaluation set, because it combines LLM skills with the discipline of measuring correctness.
Where can I find free datasets for ML projects?
Kaggle, the UCI Machine Learning Repository, GroupLens, Hugging Face Datasets and data.gov.in. Always read the licence before publishing results.
Learn this with mentors
AI and machine learning programmes with admissions open: live classes, mentors and project work.
