Picture by Creator | Canva
# Introduction
That is the second article in my newbie mission collection. If you happen to haven’t seen the primary one on Python, it’s price trying out: 5 Enjoyable Python Initiatives for Absolute Freshmen.
So, what’s generative AI or Gen AI? It’s all about creating new content material like textual content, photos, code, audio, and even video utilizing AI. Earlier than the big language and imaginative and prescient fashions period, issues had been fairly totally different. However now, with the rise of basis fashions like GPT, LLaMA, and LLaVA, every thing has shifted. You’ll be able to construct artistic instruments and interactive apps with out having to coach fashions from scratch.
I’ve picked these 5 initiatives to cowl a little bit of every thing: textual content, picture, voice, imaginative and prescient, and a few backend ideas like fine-tuning and RAG. You’ll get to check out each API-based options and native setups, and by the tip, you’ll have touched all of the constructing blocks utilized in most trendy Gen AI apps. So, Let’s get began.
# 1. Recipe Generator App (Textual content Era)
Hyperlink: Construct a Recipe Generator with React and AI: Code Meets Kitchen
We’ll begin with one thing easy and enjoyable that solely makes use of textual content era and an API key, no want for heavy setup. This app helps you to enter just a few primary particulars like substances, meal sort, delicacies desire, cooking time, and complexity. It then generates a full recipe utilizing GPT. You’ll learn to create the frontend kind, ship the info to GPT, and render the AI-generated recipe again to the person. Right here is one other superior model of similar thought: Create an AI Recipe Finder with GPT o1-preview in 1 Hour. This one has extra superior immediate engineering, GPT-4, options, ingredient substitutions, and a extra dynamic frontend.
# 2. Picture Generator App (Steady Diffusion, Native Setup)
Hyperlink: Construct a Python AI Picture Generator in 15 Minutes (Free & Native)
Sure, you possibly can generate cool photos utilizing instruments like ChatGPT, DALL·E, or Midjourney by simply typing a immediate. However what if you wish to take it a step additional and run every thing domestically with no API prices or cloud restrictions? This mission does precisely that. On this video, you’ll learn to arrange Steady Diffusion by yourself laptop. The creator retains it tremendous easy: you put in Python, clone a light-weight net UI repo, obtain the mannequin checkpoint, and run an area server. That’s it. After that, you possibly can enter textual content prompts in your browser and generate AI photos immediately, all with out web or API calls.
# 3. Medical Chatbot with Voice + Imaginative and prescient + Textual content
Hyperlink: Construct an AI Voice Assistant App utilizing Multimodal LLM Llava and Whisper
This mission isn’t particularly constructed as a medical chatbot, however the use case suits nicely. You communicate to it, it listens, it may possibly have a look at a picture (like an X-ray or doc), and it responds intelligently combining all three modes: voice, imaginative and prescient, and textual content. It’s constructed utilizing LLaVA (a multimodal vision-language mannequin) and Whisper (OpenAI’s speech-to-text mannequin) in a Gradio interface. The video walks by means of setting it up on Colab, putting in libraries, quantizing LLaVA to run in your GPU, and stitching all of it along with gTTS for audio replies.
# 4. Advantageous-Tuning Trendy LLMs
Hyperlink: Advantageous tune Gemma 3, Qwen3, Llama 4, Phi 4 and Mistral Small with Unsloth and Transformers
To date, we’ve been utilizing off-the-shelf fashions with immediate engineering. That works, however if you’d like extra management, fine-tuning is the following step. This video from Trelis Analysis is likely one of the finest on the market. Due to this fact, as an alternative of suggesting a mission that merely swaps a fine-tune mannequin, I needed you to focuse on the precise technique of fine-tuning a mannequin your self. This video reveals you the right way to fine-tune fashions like Gemma 3, Qwen3, Llama 4, Phi 4, and Mistral Small utilizing Unsloth (library for sooner, memory-efficient coaching) and Transformers. It’s lengthy (about 1.5 hours), however tremendous price it. You’ll study when fine-tuning is smart, the right way to prep datasets, run fast evals utilizing vLLM, and debug actual coaching points.
# 5. Construct Native RAG from Scratch
Hyperlink: Native Retrieval Augmented Era (RAG) from Scratch (step-by-step tutorial)
Everybody loves a superb chatbot, however most crumble when requested about stuff outdoors their coaching information. That’s the place RAG is helpful. You give your LLM a vector database of related paperwork, and it pulls context earlier than answering. The video walks you thru constructing a completely native RAG system utilizing a Colab pocket book or your individual machine. You’ll load paperwork (like a textbook PDF), cut up them into chunks, generate embeddings with a sentence-transformer mannequin, retailer them in SQLite-VSS, and join all of it to an area LLM (e.g. Llama 2 through Ollama). It’s the clearest RAG tutorial I’ve seen for rookies, and when you’ve executed this, you’ll perceive how ChatGPT plugins, AI search instruments, and inside firm chatbots actually work.
# Wrapping Up
Every of those initiatives teaches you one thing important:
Textual content → Picture → Voice → Advantageous-tuning → Retrieval
If you happen to’re simply entering into Gen AI and wish to really construct stuff, not simply play with demos, that is your blueprint. Begin from the one which excites you most. And bear in mind, it is okay to interrupt issues. That’s the way you study.
Kanwal Mehreen Kanwal is a machine studying engineer and a technical author with a profound ardour for information science and the intersection of AI with medication. She co-authored the e-book “Maximizing Productiveness with ChatGPT”. As a Google Era Scholar 2022 for APAC, she champions variety and educational excellence. She’s additionally acknowledged as a Teradata Range in Tech Scholar, Mitacs Globalink Analysis Scholar, and Harvard WeCode Scholar. Kanwal is an ardent advocate for change, having based FEMCodes to empower girls in STEM fields.
