Last month, we partnered with Lablab.ai to host the Llama 2 Hackathon with Clarifai. The Hackathon aimed to nurture creativity in the development of AI-first products using the powerful Llama 2 LLM from Meta, with any one of the 7B, 13B, and 70B parameter variants. A total of 1,765 enthusiasts comprising 227 teams registered, resulting in an impressive 51 project submissions.
With so many great entries, the task of selecting the winners proved to be a challenging one! Here are some of the standout projects.
The challenge required participants to make Llama 2 the core component of their projects while also leveraging other groundbreaking models like Stable Diffusion XL, Code Llama, GPT-4, and more.
SightCom, won the top prize with a mobile app designed for visually impaired individuals using the Flutter framework. It uses Blip 2, Color Recognition model, Optical Character Recognition model, Llama 2 7B and BARCODE-QRCODE-Reader workflow in the backend.
It offers five unique accessibility features, all seamlessly activated through voice commands:
1. Scene Description: Using Blip 2, an image captioning model, the app narrates the objects it identifies through the camera, helping users understand their surroundings.
2. Color Recognition: It identifies the colors of objects, enhancing their awareness of the visual world using Clarifai's Color Recognition Model
3. Text Recognition: Using an Optical Character Recognition model, text within the camera's view is converted into spoken words.
4. Product Reader: By scanning barcodes with the BARCODE-QRCODE-Reader Workflow, the app provides detailed information about products, empowering users to make informed choices while shopping.
5. QnA Chatbot: Powered by Llama 2 7B model, the app also offers a virtual assistant that responds to users questions.
FINGU won second place with its solution aimed at handling personal finances. Imagine FINGU as a financial expert that constantly learns from your financial interactions, behaviors, and market trends.
Here's how FINGU works:
1. User Interaction: The bot interacts with users through Telegram messages. It takes commands, prompts, and uploads of CSV files as inputs.
2. Llama 2 7B Model: To generate meaningful financial advice, it uses the Llama 2 7B model to understand user inputs and generate contextually relevant responses.
3. Role Prompt Optimization: The bot starts each conversation with a carefully designed role prompt that guides the bot's behavior. This role prompt provides useful responses, eliminating excessive randomness.
4. CSV Interaction: Users can upload CSV files, and the bot processes the content of the file using regular expressions to remove date and time formats. It then uses this processed content to formulate a conversation with the user.
5. Memory Management: It also employs LangChain's memory management that allows the bot to maintain separate conversation histories for each user. This enables the bot to recall past interactions and provide more personalized advice.
It also employs end-to-end encryption to keep information confidential and trustworthy.
SlideDeck AI, the third-place winner, is an AI-powered solution that creates presentation slide decks in just a few simple steps.
Start by describing the topic of your presentation, and based on your description, SlideDeck AI generates the content and outlines for your slides using the Llama 2 13 B Model. It then transforms your natural language content into structured JSON data using GPT 3.5. Next it creates a PowerPoint slide deck using that data.
But SlideDeck AI doesn't stop there. It also fetches helpful resources from the web on your presentation topic, giving you valuable research material, and it even creates concept art related to your topic using the Stable Diffusion XL 1.0 model.
IntelliSum is a browser-compatible extension powered by Llama 2 to enhance the web experience. It does Summarization and extraction of text from a given URL, generates AI prompts using the summarized text and can also generate images for that prompt using the Stable Diffusion model.
Definitely-Not-Jarvis another finalist is an app for instant and accurate query resolution. It uses Llama 2 at its core, understands various technical, informational, or casual queries and delivers swift and precise responses.
It can be integrated into any Slack workspace as a custom bot. Plus, it keeps learning and improving from your questions, adapting to your team's unique communication style for even better responses over time.
DefendIQ is a security solution that combines AI and IoT devices for unparalleled protection. The customizable AI adapts to unique needs, while IoT devices like cameras and drones work together to make sure everything stays safe. DefendIQ enhances human capabilities by learning from real-time responses and training on your data. It uses Clarifai's weapon detection and Meta's Llama 2 70 B models.
Provide it with input, and it generates relevant responses, making coding experiences more convenient. You can find it on the VSCode extension marketplace under the name Llama2-GPT-CodePilot.
Thanks to Lablab.ai for partnering and hosting this hackathon, and congratulations to all of the hackers that came together to make it a success. You can check out all the submitted concepts, prototypes and pitches in the Lablab.ai submission page here.
This was our first hackathon using generative AI and LLMs, and the great ideas, energy, and collaboration from all hackers involved helped us realize this to a greater level than we could've hoped. We were honestly delighted and surprised by both the number and quality of the submissions.
Want to stay in the loop for the next Clarifai Hackathon? Follow us on Twitter to stay up to date!