I'm building lilbee, the whole local AI stack in a single executable: it runs and manages the models, and it's a search engine you can talk to. It indexes your files, notes, and code, crawls web pages into your library, and every answer cites the exact file and line. It started as me trying to build an Encarta 99 out of my own stuff that I could talk to privately and terminal first, and that's still the mission.

Onboarding with local AI is painful. A typical stack is a model server, a vector database, a document ingestion pipeline, a web UI, and a Docker Compose file holding it all together. Every piece needs configuring before you can ask a single question.

lilbee is that whole stack in one self-contained executable: model manager, search engine over your own data, MCP for agents, HTTP for GUIs, and a TUI for the terminal. No containers, no networking, nothing else to install. It scales from one laptop to as many GPUs as you have.

I'm trying to democratize local AI, starting with talking to your own stuff, while building the tool I want for myself.

Terminal-first, because I built it for myself and I knew that if the architecture stayed flexible, friendlier surfaces would come easily. TUI, CLI, REST API, MCP server, and Python library, with multi-GPU placement and integrations for opencode and hermes so programmers can put local AI behind their favorite agents. lilbee is its own model manager: you don't need Ollama or LM Studio at all, and the model families it runs are tested end to end on real GPUs, the architectures behind most of the 190,000+ GGUF models on Hugging Face. Already running Ollama or LM Studio? lilbee connects to them and keeps your models.

One of those friendlier surfaces, the gateway for Q&A with local models. My favorite story here is a completely non-technical user who went from the plugin store to talking to his solar panel docs in minutes, without ever opening a terminal.

lilbee is built on a lot of open source projects, and the one I work on most is xberg.

Extraction is the core of lilbee. You can't search what you can't read, and an answer is only as good as the text that came out of the document, so a scanned PDF that turns to mush or a table that loses its columns is a bad answer waiting to happen. Good extraction is what makes good search possible. That's why I'm on the xberg open source team, and where most of my open source work goes.

In early iterations of lilbee I built my own extraction layer. When I found xberg I gutted all of that work, and a lot of it became my earliest contributions there.

Stuff I've contributed that I'm pretty proud of:

Project	Stars	What I did
xberg / alef		Semantic chunking with caller-supplied tokenizers, new formats, scanned-PDF detection, PDF page rendering for vision models, an in-process embedding plugin, reversible redaction, in-browser and ONNX-free entity detection, and native Intel Mac support. In alef, its binding generator, correctness fixes across the 16 languages it targets. All my merged PRs.
pdf_oxide		Reported and fixed a batch of text-extraction bugs in scientific PDFs: garbled math, misclassified scanned pages, split words, and crashes.
lancedb / lance		Made Lance run on older processors, with runtime SIMD dispatch and compat wheels.
GPUStack / gguf-parser-go		Reworked the multimodal VRAM estimator GPUStack sizes models with: five fixes for projectors that were over- and under-estimated (the under-estimates silently OOM), flash-attention and audio support, and unknown model types now estimated from their own metadata instead of a hardcoded list.
huggingface_hub / xet-core		Fixed a download crash in headless programs. Smoother download progress is in review.
llama-cpp-python		Some models wouldn't load because of how their chat template was written.
llama.cpp		Exposed llama-server's per-GPU memory usage on /metrics and a new GET /memory, so a program can read what each instance allocated instead of scraping logs.
bun / WebKit		Bun's Intel-Mac "baseline" build is really the AVX2 binary, so pre-AVX2 Macs crash on launch, along with everything built on bun. Building a real baseline for bun and the WebKit it bundles. Until it lands, I ship working builds.
opencode		Stopped opencode's mini mode from wiping and reprinting the whole session on every terminal resize.
candle		Fixed DeBERTa-v2 models crashing at half precision.
safetensors		Fixed large tensors failing to load on 32-bit targets like wasm.

Project	Stars	What I did
mercurius		Fixed a crash in apps with nested loaders.
screenshot-glb		Added a timeout so slow renders can finish.
MetalAPI		My first project, from college: the Metal Archives scraped and served as a REST API.
go-utils		Added unique and fixed reverse in Goldziher's Go library, years before he founded xberg and we ended up working together.
simpler-cpp-json		Automatic JSON serialization for any C++ struct, years before nlohmann/json shipped the same idea.
openapi-generator		Unity support for the C# generator.

+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
|t|o|b|i|a|s|@|l|i|l|b|e|e|.|s|h|
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+