jakecyr

I made this app to make calorie tracking less tedious. You can either:

- Take a picture of your food - Enter a brief description (ex. two eggs with slice of bacon) - Dictate what you ate - Upload existing picture - Choose previously logged meal

To track the calories and macros for those meals.

Built this simple application over a weekend that lets users sign-up, choose their interests and receive novel ideas / material about them daily. Each nugg is about a 2 minute read. Users can rate the nuggs by clicking buttons in the email to further guide the generation.

I wrote a simple library to reduce latency in voice generations from LLM chat completion streams.

This lets you generate voices from streams of text from local LLMs, such as Ollama and local TTS clients, such as Apple Say along with external clients such as Google Text-to-Speech with the same speed as privately created assistants such as OpenAI.

As each sentence end is detected, it will run TTS on it and play it out loud while the rest of the completion is being generated in the background.

Developed an open source voice assistant that integrates OpenAI's Whisper, Chat Completion and Voice Generation APIs to provide an assistant experience.

Some potential extensions could include integrating into custom hardware or adding function calling to expand the default capabilities.

The first (that I found) tokenizer for open source LLMs. It retrieves config files from HuggingFace and can encode / decode text and tokens.

Was created in an hour and can definitely use some work. Would love contributions and feedback!

Simple experiment for question answering on YouTube videos using embeddings and the top n YouTube search result transcripts.

Take a question and optionally a YouTube search query (otherwise an LLM will auto-generate one), will compile transcripts for each video result, generate an embedding index using the transcripts and then answer the question using the relevant embeddings.

Returns both a string response and a list of sources that were used for the answer.