The number one mistake beginners make when thinking of project ideas is starting with the question, “How can I use this new tech?” While this can be a fine way to learn a new tool, there is a better way.
Good project ideas start with the question, “What problem can I solve?” This not only makes for a nice story when sharing with potential employers but solving problems is how you translate technical skills into value.
The following projects all take this problem-first approach. You can take these ideas and implement them directly or (even better) use them as inspiration for solving a problem that you are personally facing.
An effective yet time-consuming part of applying for jobs is adapting your resume to different job descriptions. While automating this task would have been an advanced project a few years ago, with today’s large language models, it is as simple as an API call.
Here’s a step-by-step breakdown of how to implement such an automation.
- Create a markdown version of your resume (Note: ChatGPT can do this for you).
- Experiment with different prompt templates that take your markdown resume and a job description and output a new resume in markdown.
- Use OpenAI’s Python API to prompt GPT-4o-mini to rewrite your resume dynamically.
- Convert the markdown file to HTML and then to PDF with the markdown and pdfkit libraries, respectively.
Libraries: openai, markdown, pdfkit
While we could readily use ChatGPT for this, the upside of implementing this with Python is that we can easily scale up the process. Here’s some starter code for Step 3.
import openai
openai.api_key = "your_sk"# prompt (assuming md_resume and job_desciption have been defined)
prompt = f"""
I have a resume formatted in Markdown and a job description. \
Please adapt my resume to better align with the job requirements while \
maintaining a professional tone. Tailor my skills, experiences, and \
achievements to highlight the most relevant points for the position. \
Ensure that my resume still reflects my unique qualifications and strengths \
but emphasizes the skills and experiences that match the job description.
### Here is my resume in Markdown:
{md_resume}
### Here is the job description:
{job_desciption}
Please modify the resume to:
- Use keywords and phrases from the job description.
- Adjust the bullet points under each role to emphasize relevant skills and achievements.
- Make sure my experiences are presented in a way that matches the required qualifications.
- Maintain clarity, conciseness, and professionalism throughout.
Return the updated resume in Markdown format.
"""
# make api call
response = openai.chat.completions.create(
model="gpt-4o-mini",
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": prompt}
],
temperature = 0.25
)
# extract response
resume = response.choices[0].message.content
Note: ChatGPT is super helpful for writing short code snippets (and prompts) like this. If you get stuck, try it for Step 4.
Although I love adding technical talks to my YouTube “watch later” playlist, it might be a while before I watch them (if I ever get around to it 😅). A project that can help with this is a tool that watches the videos for me and generates concise summaries with key points.
Here’s one way to do that:
- Extract YouTube video ID from video link using regex
- Use video ID to extract transcript using youtube-transcript-api
- Experiment with different ChatGPT prompts that effectively summarize the transcript
- Use OpenAI’s Python API to automate the process
Libraries: openai, youtube-transcript-api
From a technical perspective, this is very similar to the first project. A key difference, however, is that we will need to automatically extract video transcripts, which we can feed into the LLM.
Here’s some starter code for that.
import re
from youtube_transcript_api import YouTubeTranscriptApiyoutube_url = "video link here"
# extract video ID with regex
video_id_regex = r'(?:v=|\/)([0-9A-Za-z_-]{11}).*'
match = re.search(video_id_regex, youtube_url)
if match:
return match.group(1)
else:
return None
# extract transcript
text_list = [transcript[i]['text'] for i in range(len(transcript))]
transcript_text = '\n'.join(text_list)
My watch later playlist is not the only place I hoard technical information. Another cache is my desktop, which is riddled with (118) research papers. Since manually reviewing these papers would be (very) time-consuming, let’s see how AI can help.
One could build a tool that analyzes the contents of each PDF on my desktop and organize them into folders based on topics. Text embeddings can translate each paper into a dense vector representation, from which similar articles could be clustered using a traditional machine learning algorithm like K-Means.
Here’s a more detailed breakdown:
- Read the abstract of each research article using PyMuPDF
- Use the sentence-transformers library to translate abstracts into text embeddings and store them in a Pandas DataFrame
- Use your favorite clustering algorithm from sklearn to group the embeddings based on similarity
- Create folders for each cluster and move the files into the appropriate folder.
Libraries: PyMuPDF, sentence_transformers, pandas, sklearn
The key step for this project is generating the text embeddings. Here’s a code snippet for doing that with sentence_transformers.
from sentence_transformers import SentenceTransformer# load embedding model
model = SentenceTransformer("all-MiniLM-L6-v2")
# store abstracts in a list
abstract_list = ["abstract 1", "abstract 2"]
# calculate embeddings
embeddings = model.encode(abstract_list)
A couple of months ago, I helped a company create a basic RAG system for a set of technical reports. One of the challenges with searching such reports is that key information is often presented in plots and figures rather than text.
One way to incorporate this visual information into the search process is to use a multimodal embedding model to represent text and images in a shared space.
Here’s a basic breakdown:
- Given a PDF, chunk it into sections and extract the images using PyMuPDF
- Use a multimodal embedding model (e.g. nomic-ai/nomic-embed-text-v1.5) to represent the chunks and images as dense vectors and store them in a dataframe
- Repeat for all PDFs in the knowledge base
- Given a user query, pass it through the same embedding model used for the knowledge base
- Compute the cosine similarity score between the query embedding and every item in the knowledge base
- Return top k results
Libraries: PyMuPDF, transformers, pandas, sklearn
The most important part of this project is how the PDFs are chunked. The simplest way would be to use a fixed character count with some overlap between chunks. It is also helpful to capture metadata such as filename and page number for each chunk.
Here’s some basic boilerplate code to do that (courtesy of ChatGPT). If you get stuck, try asking it to extract the images.
import fitz # PyMuPDFdef extract_text_chunks(pdf_path, chunk_size, overlap_size):
# Open the PDF file
pdf_document = fitz.open(pdf_path)
chunks = []
# Iterate through each page in the PDF
for page_num in range(len(pdf_document)):
page = pdf_document[page_num]
page_text = page.get_text()
# Split the text from the current page into chunks with overlap
start = 0
while start < len(page_text):
end = start + chunk_size
chunk = page_text[start:end]
# Store the page number with the chunk
chunks.append((page_num + 1, chunk))
# Move to the next chunk with the overlap
start += chunk_size - overlap_size
return chunks
# Parameters for extraction
pdf_path = "your_file.pdf"
chunk_size = 1000 # Size of each text chunk in characters
overlap_size = 200 # Overlap size in characters
text_chunks = extract_text_chunks_with_page_numbers(pdf_path, chunk_size, overlap_size)
# Display the chunks with page numbers
for i, (page_number, chunk) in enumerate(text_chunks):
print(f"Chunk {i + 1} (Page {page_number}):\n{chunk}\n{'-' * 50}")
Over the past year, I’ve helped almost 100 businesses and individuals build AI projects. By far, the most common project people ask about is a document question-answering system. Building on the previous project, we can implement this in a straightforward way.
If we’ve already chunked and stored our documents in a DataFrame, we can convert the multimodal search tool into a multimodal RAG system.
Here are the steps:
- Perform a search over the knowledge base (like the one created in Project 4)
- Combine user query with top k search results and pass them to a multimodal model.
- Create a simple Gradio user interface for the QA system.
Libraries: PyMuPDF, transformers, pandas, sklearn, together/openai, Gradio
Note: Llama 3.2 Vision is free until 2025 via Together AI’s API
This project essentially combines projects 2 and 4. However, it includes the essential component of a user interface. For that, we can use a dashboarding tool like Gradio, which allows us to create a chat UI with a few lines of code.
Here’s an example snippet adapted from Gradio’s doc for doing this.
import gradio as gr
import timedef generate_response(message, history):
"""
Your code for generating a response
"""
return response
demo = gr.ChatInterface(
fn=generate_response,
examples=[{"text": "Hello", "files": []}],
title="Echo Bot",
multimodal=True)
demo.launch()
Thanks to tools like ChatGPT and Cursor, it’s never been easier to build AI projects fast. Things that used to block me for hours (if not days) a few years ago can now be resolved in minutes with advanced coding assistants.
My parting advice is to use these tools to learn faster and be bold in your project choices. For projects, find problems and time-box the implementation into a weekend.
Drop your questions in the comments 🙂