Gemma

An open-source Cotypist with macOS system wide AI autocomplete

Docker Compose stack for serving a local, OpenAI-compatible LLM (vLLM on Intel XPU) on an Intel Arc Pro B60 GPU — reproducible config with operator and developer docs.

Rust + CUDA LLM inference engine for Blackwell (Tuned specifically on RTX PRO 6000, RTX 5090, B200): OpenAI-compatible (+converse and ant) serving, per-model X hardware exactness gates. NVFP4/mixed (fp8 hybrid, 4o6, etc - correctness, performance, hardware specific adapted) main quant support.

A self-improving AI agentic harness aimed at hyper-local edge use on Mac OS on llama.cpp, Gemma 4, Bonsai Ternary 1.58 Bit, and Applefoundation models

A framework for benchmarking human and large language model (LLM) text using stylometric, readability, and emotion-based features. The project supports comparative analysis across models such as GPT, LLaMA, Falcon, Gemma, OPT, and Claude to investigate patterns in machine-generated language and its similarity to human writing.

Personal Context Operating System (PCOS) — a local-first hybrid AI agent across Chrome, Android, Pixel Watch, PiecesOS, and cloud LLMs. Built on LiteRT-LM, Gemma, Chrome Built-in AI APIs, and MCP.

Queen Protocol v2.12.0 — Automated enforcement daemon (colony-watcher.sh runs every 10 min via launchd, auto-normalizes failing reports + seals stale state) + 6 operator-discipline patterns from Elev-W1 22-shard colony. 22/22 reached CONVERGE_AUDIT_PASS this session. Workers: claude/kimi/codex/gemma4-local.

W.A.R.V.I.S. — Forensic IR agent (Go orchestrator + Python MCP + Gemma 4). SANS FIND EVIL hackathon entry by WoopsFactory.

SehatGemma4 — Offline AI nutrition guard for 33M Pakistani diabetics. Snap any food → Gemma 4 analyzes risk in Urdu/English. 101 desi foods work at zero signal; vision mode works online. Built for Gemma 4 Good Hackathon.

கேளு AI — Tamil Structured Prompt Framework for Gemma 4. Helps 80M Tamil speakers prompt AI effectively using 6 structured Tamil tags. Built for Google DeepMind Gemma 4 Good Hackathon.

# AI Enabled Healthcare Chatbot AI-powered healthcare assistant built using **Spring Boot, React, MySQL, Ollama (Gemma 2B), and GPT-4**. Features include **multilingual chat, voice interaction, emergency alerts, chat history, nearby hospital suggestions, JWT authentication,** and **role-based dashboards** for patients, doctors, and admins.

An on-device, privacy-first panic attack triage agent powered by Gemma 4 & LiteRT.

Companion demo for the dev.to Gemma 4 entry on cross-model RAG drift detection (ragvitals + Gemma 4 + Claude).

Explainable Industrial AI Assistant using Gemma 4 and RAG for grounded factory diagnostics.

Ongiini AI — a free AI assistant for Namibia, and Ongiini-Eval-OW, an MT benchmark for Oshindonga and Oshikwanyama (arXiv:2609.31727)

Petit coding agent CLI local avec Ollama et Gemma 4

USB_AI

PoC codebase demonstrating Indirect Prompt Injection attacks on LLMs (Gemma 4) and mitigation techniques using Microsoft Presidio and Zero Trust architecture.

Paper: Multi-agent LLM trading for Bitcoin. 72 real calls to Gemma 7B comparing panels of 1, 3, 5 and 9 agents. 9 agents beat B&H by +2.91pp.

bulding a chatbot based on gemma finetuning

Gemma Chatbot (Ollama + LangChain + Streamlit) This project is a simple conversational chatbot built with LangChain and powered by the Gemma 2B model running on Ollama. It provides a clean Streamlit interface where users can ask questions and receive AI‑generated answers in real time.

An AI-powered clinical triage bot and physician dashboard built with MedGemma 1.5

Enclave Local is a fully offline, air-gapped AI synthesis platform for generating production-ready codebases and native PowerPoint presentations without cloud API connections.

A nostalgic, privacy-first desktop paperclip for macOS: a local Gemma model watches your screen, pops tips, runs reminders, answers questions, optionally reacts to your webcam, and scoots away from your cursor.

leverage LLM gemma3:4b to implement an text to SQL query system with two caching system: exact text match and semantic match via embeddings

Multimodal Bengali-first AI agent for street vendors, powered by Gemma 4.

An interactive AI-driven sports analytics simulation platform. It functions as a hybrid between an instant search interface and a sports management game.

Does an LLM's willingness to blackmail depend on its persona — and on whether that persona is prompted or trained into the weights? Two-part study on Gemma-3-12B: 16 prompted personas (n=100) + 10 QLoRA persona adapters, in the Lynch et al. shutdown scenario.

Context Relay saves your project context once and drops a ready-to-paste briefing into ChatGPT, Claude, Gemini, Perplexity, Grok, or Mistral — so you stop re-explaining the same project every time you switch tools. Runs on Gemma 3 via Fireworks AI.