Top Generative AI Projects to Build for Your Portfolio

Recruiters for generative AI roles see the same portfolio again and again: a chat interface over a PDF, a wrapper around an API, a screenshot of a nice answer. These show that someone followed a tutorial. They rarely show the things the job needs: getting answers right, knowing when they are wrong, and keeping costs and risks under control.

A strong project is small, built on real data and measured. The eight below each use a public dataset with a clear licence, and each comes with the number to report. Together they cover what a generative AI engineer does.

Beginner

Each of these fits into a weekend and runs on an ordinary laptop or a pay-per-use API.

1. A structured extractor. Turn support emails, invoices or job adverts into JSON with a fixed schema. Label 50 examples by hand.

• Measure: the share of valid outputs and field-level accuracy against your labels.
• Stands out with: enforced schemas, validation in code, and a list of the cases that still fail.

2. An intent router for customer messages. Route banking queries to one of 77 intents using the BANKING77 dataset, released under CC BY 4.0 (Hugging Face: PolyAI/banking77). Compare an embedding approach with an LLM prompt.

• Measure: accuracy, and accuracy at the confidence level where the system would answer on its own.
• Stands out with: a chart of answered share against accuracy, and a handover rule. See building a customer support bot.

Intermediate

These add a second component around the model, and a measurement that needs more care.

3. Document question answering with citations. Index a set of documents with clear rights, such as the Python Enhancement Proposals, which are public domain or CC0 (PEP 1). Answer questions with a quote and the section it came from.

• Measure: correct answers on 40 questions you wrote, and how often it answers questions the documents do not cover.
• Stands out with: an "I could not find that" path and a check that every quote exists in the source. See building a document Q&A chatbot.

4. A speech note-taker. Transcribe recordings with an open speech model and summarise them. LibriSpeech provides about 1,000 hours of read English speech with transcripts under CC BY 4.0 (OpenSLR: LibriSpeech).

• Measure: word error rate against the reference transcripts, with and without text normalisation.
• Stands out with: a comparison of two model sizes for accuracy and speed. See speech-to-text and text-to-speech.

5. A maths tutor that shows its working. Use GSM8K, 8.5K school maths word problems with worked solutions under the MIT licence (OpenAI: grade-school-math).

• Measure: accuracy with and without step-by-step reasoning, and the extra tokens it costs.
• Stands out with: a calculator tool for arithmetic and an analysis of the problems still missed.

Advanced

These are more engineering than modelling: state, limits, checks and comparisons against people.

6. An evaluator you can trust. Build an LLM judge and check it against people. The MT-Bench human judgments contain 3.3K expert preferences between pairs of model answers, released under CC BY 4.0 (Hugging Face: MT-Bench human judgments).

• Measure: agreement with the human majority, and how often the verdict flips when the two answers swap places.
• Stands out with: a fix for the bias you find, measured again. See evaluating LLM outputs.

7. An agent over a database. Give a model read-only tools to inspect and query a small SQL database and answer business questions.

• Measure: task success, steps per task, errors recovered, and cost per task.
• Stands out with: a read-only connection enforced by the database, a step limit, and full traces. See AI agents explained.

8. Retrieval plus a knowledge graph. Build a graph from structured metadata, such as the PEP headers, and combine it with vector search over the text.

• Measure: answers to filter and counting questions from each method, against a ground truth you compute.
• Stands out with: a router that decides which method a question needs.

How to present a project

A reviewer gives each project a few minutes. Make those minutes count:

1. One paragraph on the problem and who would use the result.
2. The headline number, with what it was measured on.
3. Three failures, with examples and what you tried.
4. Cost and speed per request.
5. Data and licence, stated plainly.
6. How to run it in a few commands.

A project that reports "93.8% of answered queries correct, with 21% handed to a person" says more than a polished interface.

Common mistakes

• Many shallow projects instead of three measured ones.
• No test set, so quality is shown by screenshots.
• Data with unclear rights, such as scraped content.
• Hiding failures. Explaining them shows judgement.
• Keys or personal data committed to a public repository.

Building chatbots, RAG systems, content generators and assistants is the subject of the Building Gen AI Applications module of the Generative AI course.

Sources

• PolyAI/banking77 dataset (Hugging Face) — PolyAI
• LibriSpeech ASR corpus (OpenSLR) — OpenSLR
• grade-school-math: the GSM8K dataset (GitHub, OpenAI) — OpenAI
• MT-Bench human judgments (Hugging Face) — LMSYS
• PEP 1: PEP Purpose and Guidelines (Python Enhancement Proposals) — Python Software Foundation