November 13, 2026, (Inside AI) — CodeNinja released a library of open reference architectures for physical AI on Thursday, publishing complete system designs for sectors including energy, oil and gas, ports, rail, manufacturing, heavy industry, aviation, and agriculture across the United States, Saudi Arabia, and Pakistan. The blueprints, released under the Creative Commons Attribution 4.0 license, include sensing requirements, model placement, object models, open-weight model and hardware choices, three-year cost projections, and the named person who approves each action.
The move targets a persistent gap in industrial AI: most organizations can deploy models but cannot inspect or adapt the design logic behind them. CodeNinja's release, branded CodeNinja Atoms, aims to shift that balance by making the design itself a public asset. The architectures are free for engineers and enterprises to study, reuse, and adapt, and the full library is available as an open dataset on Hugging Face under the name CodeNinjatools/vertical-driven-architectures, with every paper carrying its own DOI on Zenodo.
The timing reflects growing demand for sovereign AI systems that organizations can control and audit. Physical AI, which applies models to real-world operations like sensor data and machinery, often depends on proprietary designs that lock customers into vendor ecosystems. By publishing the design logic, CodeNinja is betting that openness will accelerate adoption rather than erode its commercial position.
Openness As A Path To Ownership
Many organizations use AI without retaining the design logic, model choices, object model, or operating knowledge behind it. When those elements stay external, the organization cannot independently inspect, adapt, or extend the system it depends on. CodeNinja's release puts each of those elements in the open. The reference architecture makes the design inspectable. The published object model makes the structure of the operation reusable. The model, hardware, and cost choices make every trade-off visible. The open license lets any team adapt the design to its own operation and keep what it builds.
Read: Public Funds AI Investment
"An organization cannot govern a system it cannot see. By publishing the design, the object model and the data behind each architecture, we let engineers and operators inspect how these systems work, adapt them to their own operations and build them on infrastructure they control." Muhammad Umar Bilal, Co-Founder of CodeNinja
The approach contrasts with typical enterprise AI deployments, where vendors often guard their design methodologies. CodeNinja's decision to publish cost lines and approval workflows could pressure competitors to offer greater transparency. It also gives enterprises a baseline for comparing proposals from different suppliers.
Inside AI observes that the real test will be whether published designs lead to actual deployments. Open architectures lower the barrier to entry, but physical AI projects still require domain expertise, integration work, and operational discipline. CodeNinja's own forward-deployed engineers will build systems inside customer walls, suggesting the company expects to monetize implementation even as it gives away the blueprints. That hybrid model mirrors how open-source software companies like Red Hat built businesses on freely available code.
Alongside the architectures, CodeNinja introduced Praxis, a blueprint engine for physical AI that generates complete system designs from a single operational requirement written in the operator's own words. Praxis uses a recursive learning loop where every design, benchmark run, and engineer's correction is written back to memory and applied to the next requirement, with each loop scored on a public benchmark. Engineers get a reasoning base for systems where domain expertise is scarce. The output includes sensing and model placement, the object model, hardware, three-year cost, and the approvals each action requires. Every choice arrives with the reason behind it, and a live simulation shows the operation running on the design before anything is built.
CodeNinja also published its own benchmark, the Physical AI Design Index (PADI), which gives Praxis and frontier AI the same short requirement and scores both designs against the same checks on fresh tasks across 13 industries. In Loop 22, a blind evaluation on seven unseen tasks run on 16 October, Praxis passed 48.5 percent of checks against 28.4 percent for frontier AI and led on all seven. Earlier loops ran on different tasks and show the trend rather than a like-for-like comparison. Every result is published by industry and by loop, and the scores remain far from saturated.
The benchmark results, while favorable to Praxis, come with caveats. CodeNinja designed the benchmark and the evaluation methodology, which introduces potential bias. The company acknowledges that earlier loops are not directly comparable, and the absolute scores indicate substantial room for improvement. Independent validation would strengthen the claims, but no third-party audit has been announced.
Read: Anthropic IPO AI Pricing
"We want every engineer to build on these designs and every enterprise to see exactly how a system like theirs is made. When an operator is ready for one of its own, our forward deployed engineers build it inside its walls." Muhammad Umar Bilal, Co-Founder of CodeNinja
Engineers can apply for Praxis developer access at codeatoms.ai, and organizations that want a physical AI system built for their own operation can contact CodeNinja. The company has not disclosed pricing for its implementation services or the terms of developer access.
The release lands as governments and enterprises increasingly seek AI systems that can be deployed on domestic infrastructure. Saudi Arabia and Pakistan, both named in the architecture coverage, have active initiatives to build local AI capabilities. By providing open designs tailored to these regions, CodeNinja positions itself as a partner for sovereign AI projects. The long-term impact will depend on whether the open library attracts a community of contributors and whether CodeNinja can convert free users into paying customers for its engineering services.
Shamil has a low tolerance for AI that only works in controlled demos. He covers agentic systems and generative AI with the skepticism of someone who has seen too many “revolutionary” tools quietly disappear six months after launch.
