MAZHARMIK/python-ai-agent-ollama — A bare-bones local AI agent you can read line by line
A Python AI agent built from scratch that runs on Ollama with no framework, showing exactly how tool calls, memory, and multi-step loops work.
Open-Weight Models
Practical reporting on AI systems, tools, and engineering execution.
Git repos
prabanta-dev/janas: C-based inference runtime for running 48 GB LLMs on 32 GB laptops
Janas is a C-based inference engine that runs large language models exceeding system memory on consumer hardware, using GPU and NPU acceleration alongside CPU execution.
Speech Studio Teardown: Rust-Based Local Voice Synthesis with Built-In MCP
Speech Studio packs offline text-to-speech, transcription, and an MCP server into a native Windows executable built with Rust and CUDA without relying on Python runtimes.
Teardown: Fine-Tuning Qwen2.5-Coder-7B into an Execution-Verified Local Tutor
A practical blueprint for fine-tuning Qwen2.5-Coder-7B via QLoRA on free Kaggle GPUs to run an offline coding agent within an 8 GB RAM footprint.
Converting Hugging Face Models to On-Device .litertlm Artifacts
An automated conversion and verification toolchain that translates open-weight Hugging Face LLMs and VLMs into deterministic .litertlm bundles for Google's LiteRT-LM runtime across mobile and desktop platforms.
Atomic Agent Combines Local-First Execution with Hybrid Model Fusion
Atomic Agent runs local autonomous workflows using optimized llama.cpp quantizations, supporting split planning and execution models alongside direct imports from existing agent toolchains.
TensorSharp brings local GGUF inference and agent skills to .NET 10
TensorSharp gives .NET teams local GGUF inference, compatible APIs, accelerator backends, multimodal workflows, and agent tooling subject to model-specific validation.
goinfer puts constrained local LLM inference inside Go services
A pure-Go runtime lets Go applications embed local models, enforce JSON Schema outputs, and optionally expose OpenAI- and Anthropic-compatible APIs.
cliproxy-rs: Rust proxy bridges Codex and Claude to DeepSeek, GLM, and OpenRouter
A Rust rewrite of CLIProxyAPI that lets Codex CLI and Claude Code call multiple providers through one local endpoint with API translation and failover.
comanda: A CLI-native orchestrator for multi-agent AI workflows with quality gates
Orchestrates multi-agent AI workflows via YAML-defined pipelines with quality gates, enabling iterative code improvement using existing tools like Claude Code and Codex.