AI platform
In developmentCrosswind AI
Run your own models, on your own hardware, with no data leaving the building.
Overview
A suite of compiled AI tooling built to run on your own hardware: train domain-specialised LoRA adapters, serve them behind an OpenAI-compatible API, generate images and meshes locally, and feed industrial telemetry in through SCADA. Pure Rust and C++ throughout — no Python runtime anywhere in the stack.
What it does
- Specialised LLM serving
- Multi-domain LoRA adapters, hot-swappable without a restart, behind a type-safe OpenAI-compatible REST API. Training on Burn, inference on Candle, API on Rocket.
- GPU LoRA training
- A C++ fine-tuning tool on ggml/llama.cpp with Vulkan, CUDA and ROCm backends, exporting GGUF adapters that the inference engine loads directly.
- Local image generation
- Stable Diffusion inference in C++ via stable-diffusion.cpp — a generation pipeline that never leaves your machine.
- Mesh generation
- 3D asset generation on a ggml backend, sharing the same compiled, dependency-light approach as the rest of the suite.
- SCADA integration
- A Rust service bridging industrial telemetry into the platform, so models work against live plant data.
- Field mapping and RAG
- AI-assisted schema mapping using embedding similarity with optional LLM fallback, plus retrieval over your own corpus.
Built with
- Rust
- C++
- Burn
- Candle
- Rocket
- ggml
- Vulkan / CUDA / ROCm
Worth knowing
Why compiled, and why self-hosted
No Python dependency chain to reproduce, no interpreter in production, and no customer data leaving the building. It deploys as binaries and runs on the GPU you already own.
Interested in Crosswind AI?
Ask about availability, licensing, or having something similar built for your own domain.
Get in touchMore work
Fuel Planner
Aviation app
Cross-country fuel planning that answers the question EFBs can't.
Genus
Game
Puzzles you solve by reasoning, not by memorising — where the surface itself is the problem.