Jul 16, 2026
Multi-agent RAG platform delivering lifelong personalized tutoring via CLI and web interfaces
Implemented in Python with Next.js frontend, it combines LightRAG, GraphRAG, and partner agents that maintain private memory and skills while supporting document ingestion, quizzes, and live tool calling. The design applies established agent and retrieval patterns to continuous learner modeling, limiting broad adoption to edtech teams building custom tutoring workflows.
WebKit compiled to WebAssembly running a full browser engine inside a browser tab
WebCore and JavaScriptCore are cross-compiled via a custom Emscripten port with Skia rendering to an OffscreenCanvas or direct WebGL context and Wisp replacing the curl transport layer for all network I/O. The resulting single-process engine enables sandboxed browsing sessions or embedded browser components for specialized developer tooling and research environments rather than replacing conventional browser usage.
Rust TUI harness providing fullscreen mouse-driven interaction for extensible coding agents
The harness implements an agent execution loop inside a terminal interface that accepts mouse events and supports runtime extension points for new tools or models. It follows established patterns for agent orchestration and TUI frameworks rather than introducing new primitives, limiting its audience to teams already building LLM coding workflows.
Python toolkit for reproducible sparse and dense retrieval research using Lucene and Faiss
Pyserini wraps Anserini for Lucene-based sparse retrieval and Faiss for dense vector search while supplying prebuilt indexes and evaluation scripts for standard test collections such as MS MARCO and BEIR. The approach standardizes first-stage retrieval pipelines in multi-stage ranking but targets the specialized information retrieval community rather than general machine learning practitioners.
Native Swift game streaming client with HDR pipeline for Sunshine and Moonlight hosts on Apple Silicon
It implements the GameStream protocol end-to-end using VideoToolbox decoding into AVSampleBufferDisplayLayer plus an AVAudioEngine pipeline, with optional privileged Wi-Fi helper to disable AWDL. The native-first Swift approach improves macOS integration and 10-bit HDR fidelity over Qt ports but targets a narrow audience of Apple Silicon gamers rather than becoming a foundational primitive.
Browser playground for running Gemma models entirely on-device with WebGPU
It loads models from Hugging Face via Transformers.js, executes inference in the browser's GPU context, and additionally exposes Chrome's built-in Gemini Nano APIs when the required flags are set. The approach follows established patterns for client-side LLM execution rather than introducing new primitives, limiting its audience to developers already exploring on-device multimodal inference.