I’m noobli, an independent builder focused on AI applications. I work with agents to develop software, connect tools, and verify results for problems I encounter. From video production to full-stack applications, I care about getting an idea to work—and learning how to build the next one better.
PROJECTS / IDEA TO USE
This site runs on a tool I built and published.
I led the development and publication of Viselora, a runtime connecting DOM content and WebGL visuals through public APIs. This website puts it to use. Another project, AXMORF Studio, organizes video production into creative tasks for agents and verifiable steps for software.
FULL STACK / DELIVERY
From pages and APIs to data and deployment.
For Luju Living, I worked with agents on room selection, quotes, bookings, and check-in, deploying the service and database to an Ubuntu demo environment. I also explore local vision measurement, document Q&A, and development-journal pipelines. Each project records its progress and verification limits.
EXPLORATION / JUDGMENT
Understand the problem, then choose the tools.
I discover tools through conversations with AI, official documentation, and open-source experiments. I compare approaches and try small tasks before committing, changing direction when the results call for it. I don’t claim mastery of every tool; I learn the parts that matter and keep revising my judgment.
COLLABORATION / INDEPENDENT JUDGMENT
People express intent. Agents organize the work.
In June, I designed and tried a single entry agent coordinating a remote team in my Mac-to-Ubuntu workflow: discuss the goal locally, then let that agent coordinate execution. Dots later echoed this direction. I want a clear human entry point as the execution system evolves, with people retaining control over goals, boundaries, and final acceptance.
For me, capability in the AI era shows in the questions we ask, the work we deliver, and the methods we keep improving.
Axioms
AI rewrites the premises
When an agent can carry a task from interpreting a request through execution and verification, the change reaches beyond coding speed. I find myself reconsidering how personal capability becomes visible: the problems we choose, the tools we organize, the work we deliver, and the reasons we trust it. I want my résumé to show that process, with projects as evidence.
One human entry point, an evolving execution system
In June 2026, I independently designed and tried a layered workflow for my Mac-to-Ubuntu work: discuss goals with the local m-hermes, let it prepare the task, then have it coordinate a remote team through that team's Kanban process. My June 2 conversations already distinguished the entry agent from the remote coordinator and asked agents to handle handoffs, reducing manual context copying and dispatch. OpenAI introduced Dots on September 29 that year. Its documentation describes taking on goals, coordinating background agents, delegating Work or Codex tasks, and returning results. Those collaboration choices echo my earlier judgment. I want models, tools, and execution teams to evolve while people retain a clear entry point, their intent, and decision authority. Task boundaries, acceptance criteria, and execution evidence make this collaboration accountable.
The more I talk with AI, the more I want to distinguish fluent expression from understanding a problem. Language helps organize my thinking, but it can also conceal omissions and misunderstandings. Could cognition include forms that language cannot fully express? I keep that question open. In practical work, I test convincing explanations against sources, code, and actual results.
Giving agents more work does not remove the need for judgment. My publishing automation uploads media, fills fields, and reads back the page state, leaving final publication to a person. The same distinction matters in code: an operation executed, tests passed, and a release deployed are different states. Authorization, acceptance, and external commitments need an accountable owner.
My own experience has limits. Prescribing every step too early can confine AI to the answer I already have in mind. I prefer to explain the goal, context, and constraints, then ask it to research practices, compare options, and expose blind spots. I weigh the sources and make a decision. Even a best practice still has to prove useful for the problem at hand.
I tried several browser-automation approaches and currently use Ego Lite because it fits my work. My 3D asset experiments also took me from Blender MCP to Hunyuan image-to-3D. I value knowing which part of a problem a tool can solve and when a replacement is worth trying. Small experiments and continued use must keep that judgment current; a list of bookmarks is only a starting point.
I use separate accounts, a dedicated Mac user environment, and an Ubuntu machine for agent work, connecting remotely through Tailscale. This helps distinguish project resources, login state, and everyday personal data, while reducing interference from old configuration. It cannot guarantee that an agent will avoid mistakes, but it makes the resources it uses and the actions it takes easier to trace.
Correcting the same problem repeatedly taught me to value Skills, engineering conventions, and maintained documentation. I record entry points, responsibilities, steps, and acceptance conditions, then put repeatable checks into scripts and tests. When implementation changes, documentation must follow and old plans must be archived. That gives an agent in the next conversation a reliable place to continue.
A standalone bilingual personal site built with the Next.js App Router. It combines a profile, essays, project showcases, public channels, and a daily journal while using public Viselora packages for DOM, scrolling, and WebGL experiences.
A composable WebGL scroll-effects toolkit for React, Next.js, and agentic frontend workflows. It combines a layered core runtime, built-in effects, React triggers, viewport-distance lifecycle scheduling, and hooks for host-managed asset prefetching.
A Chinese-first spatial, AI-native personal homepage that uses long-form scrolling, a fixed WebGL canvas, and a single recursive monolith to express five stages: observation, causality, recursion, self-reference, and reconstruction.
An AI-first video workspace built with TypeScript, Next.js, and Remotion. Its primary route produces dedicated videos from real topics using research, evidence assets, TTS-first timing, and custom compositions; an editable web generator remains a secondary productization path.
A local multi-agent workspace that connects Hermes planning and review with Codex implementation and testing through structured handoff artifacts, while reserving direction, constraints, approvals, and final decisions for the human user.
A local video production workspace built on Remotion. A coding agent organizes scenes, narration and covers, while a defined production process and validation deliver a video, two covers and a publishing manifest. Valid artifacts can be reused, and revisions are developed in isolation.
A DOM-first WebGL runtime for React and browser applications. It preserves web layout and interaction semantics, connects text, media and 3D objects through public declarations, and manages rendering, resources, input and lifecycle in one place.
A syringe measurement application that runs locally on the CPU. Detection, orientation, range markers and plunger position produce continuous volume readings, live charts and CSV records controlled by the user. This case study follows a vision prototype through desktop interaction, failure handling and Windows distribution.
A development journaling pipeline built with the Python standard library. It brings together Hermes, Codex and optional OpenCode sessions, produces structured Chinese journals by date, rebuilds a technology inventory and timeline from those journals, and supplies selected public records to this website.