Email me
← V Naman Mehta

Projects

Things I have built to learn how they work. 11 so far, most recently updated first. Each one links to its source.

  1. Nova

    A tray-resident desktop assistant, summoned with a hotkey, meant to act on the machine through an LLM tool-calling loop. Risky actions will wait for the user to confirm.

    Electron, React, TypeScript

    In progress. The shell, hotkey and typed IPC bridge are built. The agent loop is not wired up yet.

  2. envanalyzer

    Traces which environment variables a JavaScript or TypeScript project reads, with a confidence score for each result. Published as two npm packages.

    TypeScript, Node.js

    Public, MIT licence

  3. Semantic Screenshot Daemon

    Runs in the background, watches for new screenshots, has Gemini Vision describe each one, and an organisation agent picks the folder and file name. A vector store of past screenshots keeps its choices consistent.

    Python, Gemini Vision API, FAISS

    Public

  4. Small language model

    A GPT-style transformer written from scratch: a custom BPE tokenizer, 8 layers and about 9 million parameters, trained on TinyStories.

    Python

    Public. Validation loss 1.35, perplexity 3.84.

  5. Mini Browser

    A lightweight tabbed web browser with navigation controls and a combined address and search bar.

    Electron, React 19, TypeScript, Vite

    Public

  6. Multimodal document Q&A

    Ask questions about a PDF and get answers grounded in both its text and its images, using CLIP, a FAISS index and the Gemini API.

    Python, Flask, Next.js, TypeScript, LangChain, FAISS

    Public

  7. Restaurant review RAG API

    Answers questions about a pizza restaurant from its customer reviews, and indexes new reviews the moment they are added. The models run locally.

    FastAPI, LangChain, ChromaDB, Ollama

    Public

  8. Code Reviewer

    Paste a code snippet and get an AI review of it.

    Express, React, Vite, Gemini API

    Public

  9. News summarizer and sentiment analyzer

    Summarises news articles and analyses their sentiment, with a FastAPI backend and a Next.js front end.

    FastAPI, Next.js, TypeScript

    Public

  10. Virtual camera controller

    Segments a person in the webcam feed with YOLOv8 and blurs, blacks out or replaces the background, then streams the result to a virtual camera. A small web page controls it.

    Python, YOLOv8, OBS virtual camera

    Public

  11. AI Flappy Bird

    A neural network learns to play Flappy Bird, evolved over generations with the NEAT algorithm.

    Python, Pygame, NEAT-Python

    Public

Loading