Light Ark Technologies
Models are public infrastructure.
The harness is sovereignty.
The model everyone can rent. The layer wrapped around it — memory, rules, tools, judgment — is what makes an AI yours. Light Ark Technologies builds that layer, and hands it over already built.
The Ownership Gap
Owning your AI today means being able to build a harness around a model: memory that survives the session, rules it cannot talk its way out of, tools it is allowed to touch, and a way to tell whether it did the job. That is engineering work. Most people will never do it — not because they lack ideas, but because they have no engineers.
We call that distance the Ownership Gap. Our answer is not to teach everyone to build a harness. It is to give them one that is already built — the software equivalent of handing someone a finished house instead of a pallet of bricks.
ArcOS
A model-agnostic harness
ArcOS is built on one principle: perception belongs to the language model, while control, persona, memory and rules belong in hard-coded logic outside it. A five-stage pipeline (perception → strategy → assembly → generation → memory) keeps an AI consistently in character and gives it durable long-term memory — on any backbone (Gemma, Llama, Qwen). Swapping the model is a configuration change, not a rewrite.
Rules that outlive the model
Behavior is governed by written, versioned rule documents that the pipeline enforces in code rather than by asking the model nicely. When something goes wrong, the fix lands in the rules and stays there — including after the underlying model is replaced.
In production today
Our first product — a 24/7 AI companion for everyday support and reflection — runs on ArcOS and is live on the web and on desktop (macOS, Windows, Linux), with a native iOS app in development. It runs on a dedicated NVIDIA H200 we own and operate.
Open source
founder-coding-skills
The working methods we use to run an AI engineering team without engineers, packaged as Claude Code skills and released under the MIT license. Thirteen skills in two layers: how a non-programmer founder runs the team, and how the AI disciplines its own engineering. Shipped with the case studies behind them — including the postmortem of the week we burned our entire quota in three days, published with the real numbers.
Research
The Institutional-Rule Prefrontal Layer
Governing models from outside the weights: how an institution’s rules, cases and corrections can live in a durable layer above the model, so that experience accumulates across model generations instead of being reset by them.
Paper (GitHub) Essay version DOI: pending
The Ownership Gap
How people who cannot build a harness can still own their AI — the case for shipping a pre-built sovereign harness rather than more parts. Companion paper to the Prefrontal Layer.
Paper (GitHub) Essay version DOI: pending
Company
Light Ark Technologies, Inc. is a California C-Corporation (formed 2025) building applied AI for real-world deployments. Founded and led by Haijun Wen. Based in Chino Hills, in the Greater Los Angeles area.
We work on one question in three places: the engine for people who have no engineers, the open-source pack for founders who never wrote code, and the published research for anyone who has to govern an AI they did not build.