Everything that runs on the metal: local model serving behind one OpenAI-compatible endpoint, agents in production, air-gapped reference deployments, and the open retrieval toolchain. Private AI that gives up nothing, proven on our own pair before it reaches yours.
PRIVATE BY ARCHITECTUREEverything that runs on the metal: local model serving behind one endpoint, agents in production, air-gapped reference deployments, and the open retrieval toolchain. Private AI that gives up nothing.
The flagship entry is SparkDuet, the open source operating layer for the two-Spark pair. Its page is the template for every entry that follows.
The open source operating layer for a two-Spark pair: run DeepSeek, Qwen, and your own fine-tunes behind one OpenAI-compatible endpoint with four serving lanes and honest numbers. Pick the topology per model and per workload, switch in one command, revert in one command. MIT licensed.
Every client speaks the OpenAI API to one address. Behind it, SparkDuet routes to four serving lanes: one big model across both nodes, one model per node, many small models, or a serve-and-train split. The lane is a per-model decision, switched in one command and reverted in one.
The same stack runs in the cloud and fully air-gapped. The manufacturing reference deployment runs six local models, offline retrieval over 765,000 chunks, and 57 tools with zero external dependencies: the whole capability, not a reduced mode.
Models, retrieval, and agents running on hardware you control: the Lab's own two-Spark pair in daily operation, and client deployments up to fully air-gapped. The same stack runs in the cloud and offline; environments where data cannot leave the building get the whole capability, not a reduced mode.
DeepSeek and Qwen family models plus the Lab's own fine-tunes, served with vLLM behind one OpenAI-compatible endpoint through SparkDuet's four lanes. The lane layout is chosen per model and per workload, switched in one command.
Yes. The reference case is a manufacturing deployment with six local LLMs of about 58 GB, a 765,000-chunk offline retrieval corpus, and 57 callable tools, entirely behind the plant firewall with zero external dependencies.
The Lab does, around the clock: the Hermes agent series documents the full path from install to production on local vLLM-served models, including tool permissions, security hardening, a Telegram gateway, scheduling, and memory, in eighteen posts.
The open toolchain: DocPrep-AI prepares enterprise documents for retrieval entirely in the browser, Weaviate Local UI makes local vector data inspectable and queryable, and the retrieval engineering posts document the failure modes in between. All open source under github.com/zorost.
We don't pitch slide decks. We show you what we've already built in your domain, then engineer what your mission requires.
