Projects I'd like to showcase. A full list can be found here.
I also post on Works · Læra Labs.
An app to compare responses from multiple local LLMs with customizable temperature, system prompts, and CSV/JSON response exports.
Utilized the Ollama Python library and API for local model inference, automatic model detection, agentic response evaluation, and in-app model management
A project I worked on at Ignition Hacks 2024. Sty.ly generates outfit recommendations based on real-time weather data.
I used llama3.1 to generate a custom dataset of 30k entries for outfit recommendations based on weather combined over rain, snow and other conditions.
I then trained llama3.1 8B on this dataset using Unsloth @ 375 steps and quantizing it to Q4_K_M.
This trained model is run locally to get an outfit recommendation from real-time weather data which is retrieved via a free API.
The outfit recommendation is then sent to a locally hosted venv instance of Stable Diffusion via A1111SDK, which generates a model wearing the generated outift.
We then encode the generated model into an image and then into Base64, sending it to our database. After the database receives it, it is displayed on the frontend.
A project I worked on at HackThe6ix 2024. devDucky is an integrated IDE that does more than hold your code.
I used Unsloth to fine-tune and quantize three different models (llama3.1 @ 375 steps, tinyllama @ 1 epoch, phi3 @ 375 steps), all quantized to Q4_K_M.
We settled on phi3 due to hardware constraints.
I used the Ollama Python library to run inference on the models.
App is built with LangChain and uses Ollama for model inference. Llama3.1 8B is, by default, the main model. The RAG approach is achieved via LangChain and uses GPT4ALL embedding. It defaults to .md files for its intended use with Obsidian.md.