NOUMENAX

Keanu Shirani / Independent engineering

Intelligence made sovereign.

BUILDING INTELLIGENCE FOR THE PHYSICAL WORLD.

Keanu Shirani — Robotics, AI & Systems Engineering

Explore the engineeringThe philosophy

Interactive robotics simulation · Real-time 3D

Autonomous kitting and inspection in the browser: rule-based simulated perception, analytic IK, checked paths. Not footage of hardware, HOUSE, or the Ethernovia project.

Open simulation labEnter the archive

From the concept to the engineering

Flagship / Collaborative initiative

Planned Development

NETWORKED PHYSICAL INTELLIGENCE

Ethernovia / Robotics / Semiconductor Networking / Physical AI

Developing articulated robotic arms to demonstrate how high-speed networking connects machine perception, intelligent computing, and coordinated physical action.

Six-Month Engineering Initiative / Engineering intern

Full project case study

Existing technical software / Public source

In Development

HOUSE

CPU-first simulation & AI-agent evaluation

A working research prototype for repeatable household-agent experiments, with documented implementation, checks, and known limits.

Human-led / AI-assisted

Explore HOUSEInspect the source
01Evidence

What exists today.

Engineering index

HOUSE is the work that can be inspected right now: public source, documented scope, and stated limits.

Open the HOUSE episode lab
Conceptual symbolic environment diagram, not a simulator screenshotXZY
Symbolic environment study002
Conceptual illustration / Not project evidence
002 SimulationIn Development

HOUSE — CPU-First Robotics Simulation & AI-Agent Evaluation Framework

My roleCreator, Systems Architect & Lead Developer

Research prototype under active development

I conceived and architected HOUSE, a CPU-first household robotics simulation and AI-agent evaluation framework. I led its technical direction, diagnosed bugs, evaluated solutions, tested and integrated components, and used AI coding agents extensively for implementation.

PythonCPU-firstProcedural generationGeometric perceptionAgent evaluation
Engineering case studyView on GitHub

HOUSE / Documented scope

Coverage & limitations
Cataloged requirements
2,720
Authored scenarios
242
Scenario verification
242 / 242

Repository-reported checks of authored scenarios in their declared perception modes across five documented seeds; not independently rerun here. Cataloged requirements are not an executable-task count.

Read the evidence
02Engineering decisions

Why it was built this way.

A notebook of the choices behind the work: the problem, what constrained it, what I decided, and what is still unknown. Fields that haven't been written up yet say so.

DR-001HOUSEStay on Python's standard libraryOpen record +
Problem
Choose the runtime foundation for a household simulation that has to run offline on an ordinary computer.
Constraints
The original contract excluded physics dependencies. Windows with Python 3.14 had no reliable physics wheels, and the specification named no engine.
Alternatives considered
Not yet documented.
Decision
Stay on standard-library Python and label fidelity explicitly: S (symbolic), G (grid geometry: reach, clearance, paths, fit, stacking, stairs) and P-approx (lumped physical models). Full rigid-body, deformable or fluid simulation is not claimed.
Reasoning
Approximate physics must never be relabeled as symbolic success. G/P fidelity is treated as separate, future work.
Implementation
The project runs entirely on Python's standard library. No GPU, Isaac Lab or API key is used.
Results / current status
Runtime tested on Linux with Python 3.12.14; source syntax checked against Python 3.9. Windows and a Python 3.9 interpreter were not available for testing.
Lessons learned
Not yet documented.
Source: HOUSE / DECISIONS.md
DR-002HOUSESeparate catalog metadata from executable coverageOpen record +
Problem
A catalog of 2,719 requirements makes it easy to overstate how much actually works.
Constraints
Not yet documented.
Alternatives considered
Not yet documented.
Decision
No keyword-only or generic task_done switch may count as a solution. DONE means a verified, bounded S-level scenario for a source entry, not real-world task competence.
Reasoning
Each supported family has to be judged on observable state changes and protected invariants, so unsupported behavior stays visible instead of hidden.
Implementation
One authored, data-only spec per catalog entry (tasks/authored), each verified separately. Refused unsafe, private or unauthorized attempts fail a task by default, so an agent cannot probe its way through guards.
Results / current status
Only entries with verified executable scenarios are marked DONE in coverage.csv. The full brief remains incomplete, and the final report states it.
Lessons learned
Not yet documented.
Source: HOUSE / DECISIONS.md
DR-003HOUSEEnforce layout realism with invariantsOpen record +
Problem
Generated houses must stay navigable and realistic as many domains add furniture and objects.
Constraints
Not yet documented.
Alternatives considered
Not yet documented.
Decision
Enforce layout realism with invariants on generated houses instead of relying on each furnisher to be careful.
Reasoning
A placement may never make access worse: doorways stay mutually reachable, everything that was reachable stays reachable, and the robot can always leave its dock.
Implementation
Loose items never sit on the only hazard-free route. A final repair pass and an audit tool keep this true as domains evolve. Operable fixtures (windows, switches, valves, panels) are protected; things real homes make unreachable, like an outlet behind a fridge, are allowed.
Results / current status
Not yet documented.
Lessons learned
Not yet documented.
Source: HOUSE / DECISIONS.md

02Method

Engineering in the age of AI.

AI coding agents wrote a substantial share of HOUSE. That is stated plainly on the project page. What they did not do is decide what HOUSE should be, which tradeoffs to accept, or whether a result was actually correct.

I treat agents as implementation accelerators. Accountability for the architecture, the tests, and every claim on this site stays with me.

How it worked on HOUSE
  1. 01Define the problem
  2. 02Architect the system
  3. 03Evaluate tradeoffs
  4. 04Direct implementationAgents assist
  5. 05Test assumptions
  6. 06Debug failures
  7. 07Integrate components
  8. 08Verify outcomes

Human-led at every step except where marked

03Research directions

The questions worth building for.

Research archive

Open questions, not claims of solved problems.

01Multi-arm robotics

Coordination without conflict.

How can independent arms share space and make useful progress without obstructing each other?

Related engineering

02Computational accessibility

Useful simulation on ordinary hardware.

What can we evaluate before depending on an expensive GPU or a physical robot?

Related engineering

03Observation & verification

Evaluation an agent cannot bluff.

How do we know a policy solved the task without leaking the answer through its observations?

Related engineering

03The long-term direction

This is not just a portfolio of what I have built. It is the foundation of what I intend to create.

The person behind NOUMENAX

Keanu Shirani

Intelligence made sovereign.

I’m driven to architect, build, and advance systems that expand what technology can accomplish in the physical world. AI, robotics, and simulation are not just technologies I follow. They are the means through which I intend to turn that ambition into working systems.

The engineerThe philosophyExperience & leadership