We build deterministic engines and assistants: same input, same output, every answer traceable to a rule or a source. When it doesn't know, it stops and asks a person. Runs on your hardware. No cloud, no GPU.
Everything below exists and runs. Two are public. One is in delivery for a client. The rest run on-premise and can be shown live. Nothing here is a mockup: the window below is a real session, captured 2026-09-08.
Custom family illustration on Shopify. The engine assembles the artists' own approved artwork by the rules the artists defined and outputs print-ready files. When two names collide or a piece needs an adjustment, it flags "review" instead of guessing. No generative AI in the composition.
Four layers on every CV: completeness, positive signals, cognitive profile, red flags. A holistic match against the role. Runs on the recruiter's own Mac; nothing leaves the machine. Signed licenses and a merchant of record for billing.
A legal assistant that answers only from the firm's own documents, quotes the passage verbatim, and stops when the answer isn't in the source. Deterministic retrieval. 100% on-premise, packaged as a macOS app.
A real clinical system: writable on-premise clinical database, patient records, appointments, population recall, and preventive guidance cited to USPSTF. Bilingual, built for US primary care.
Simulated robots run by the deterministic core in a live web demo: warehouse, restaurant, battle, procedures. Every decision observable, step by step. Open the demo.
On-device, no cloud. Memory in five hemispheres: lexical, graph, procedural, vault, vector. Bilingual. It learns to speak from use. Native Android, running on a mid-range tablet.
Zalvyum is the memory-first deterministic core under every body above: routing, memory, composition, and a judge that decides whether an answer is grounded. The language model, when there is one, is a replaceable part that only phrases the answer. It never decides what's true. Measured 2026-09-08 on a laptop, same ten questions: 100 MB and no GPU, against 5 GB and the GPU for a 7B model.
Read the Zalvyum thesis →Two ways we build for clients. Both start by measuring your real data before a line of code is written, and both ship with their own automated checks, so you know it worked without checking by hand.
A short call to find the repeatable half of the problem. A measured inventory of your real data. A fixed-scope build that runs on your hardware and replaces the decision your team makes by hand, deterministically.
Legal, clinical, or any regulated domain: an assistant that answers only from your sources, quotes them verbatim, and stops when the answer isn't there. Nothing leaves your machines.
Zymbiotech designs, builds, and ships the core, the bodies, and the judges that check them. No resold models, no assembled demos.
Founder of Zymbiotech and architect of the Zalvyum core. Leads every engagement personally: the discovery, the measured inventory, the build, and the judges that prove it works. The differentiator isn't the AI. It's that what Zymbiotech ships doesn't guess.
LinkedIn → cospraxFor an engine built to spec, an on-premise assistant, or to see any of the systems above running live — write.