Takayoshi Nakamura GomezT. Nakamura Gomez

Personal lab · Singapore and Japan

I turn problems into working AI tools, with a team of agents.

It started with an online course on AI agents in July 2025. Seven months later, my own AI assistant was running day and night. Since then, when I hit a problem worth solving, I build the fix with a team of AI agents. These are the results.

Follow on XConnect on LinkedIn

The journey · July 2025 to now

Six builds, oldest first.

Six tools I built and use myself, and what each one taught me.

  1. A 30-second loop in painted anime style. Meet Sky, on when my laptop is off: Sky, an AI assistant drawn as a calm professional with a sky-blue bob, works at the night desk while the laptop is shut. She keeps her list (checks all good, backup restored, 2 notes need your OK). In the shared vault she files notes and offers one for approval: she suggests, I decide. At sunrise she sends a good-morning message. In the evening she lines up picks: a quiet film, new places for dinner, something to learn. Screens are illustrated with dummy content; scenes are AI-generated.

    Project 01 / 06 · Feb 2026 · AI assistant

    Sky is my personal AI assistant, running on a server I control.

    Always oneven when my laptop is closed

    The problem

    I wanted an assistant that's always there and remembers what we talked about.

    What I built

    An AI I message in chat. It runs routine jobs on a schedule and keeps a list of what's waiting on me.

    Learned →Running your own AI means owning the boring parts too, like backups that really restore.

    Hermes · My own server · Claude Code · Codex

  2. A 40-second illustrated shared-memory loop. One notebook sits between four glowing glass globes for Claude, Codex, Grok and Cursor, and an open pedestal where blue-haired Sky gently hovers and reads a note. Small custom emblems and discreet names identify the peers. Claude suggests the dummy note Use plain words; I choose to keep it. Codex and Sky read paper copies while the original stays in the vault. Cursor joins and reads the same kept note. Grok steps away; its connection dims while the note remains. Unkept suggestions fade separately. The vault's real graph, with labels hidden and motion added, closes the film: Agents change. The memory stays. The workflow uses dummy notes; archive scenes are AI-generated.

    Project 02 / 06 · Jun 2026 · AI memory

    Sky Vault gives my AI agents one shared memory, even as the team changes.

    One memoryfor Claude, Codex, Grok, Cursor and Sky, and I decide what stays in it

    The problem

    Each AI was learning the same things about me and my projects, separately.

    What I built

    A shared notebook of plain notes. The agents suggest additions; only the ones I approve stay.

    Learned →AI remembers better when it suggests and a person decides.

    Obsidian · Markdown · Git · Sky · Claude · Codex · Grok · Cursor

  3. A 19-second loop. Every lap, something worth hearing, over an empty pool lane at dawn. In SwimCast, an episode gets five stars (Rated), the app recommends what to hear next with a reason for each pick (Recommended), and an AI summary sums up an episode before it is added to the swim (Summed up). One click on Sync to device: the episodes are sped up, evened out and copied to the SWIM PRO headphones, and the app confirms Synced 5. A swimmer puts the headphones on and glides underwater while Ep. 112, Breathing on both sides, plays. Pool scenes are AI-generated; screens are recreated with dummy shows.

    Project 03 / 06 · Jun–Aug 2026 · Desktop app

    SwimCast loads a swim's worth of podcasts onto my waterproof headphones.

    One clickfrom podcast queue to headphones, ready for the pool

    The problem

    Loading podcasts onto swim headphones was the worst part of swimming with them.

    What I built

    My own version of an open-source podcast app that does the loading for me, sped up and evened out in volume.

    Learned →AI can help you ship a real app, but the risky parts still deserve a second look.

    Tauri · React · TypeScript · Claude Code · Codex

  4. A 29-second loop in painted anime style, set in Kansai. Quiet, until a place fits: a phone on a morning table by a window over a hillside street buzzes, and Sky messages: JUMP, House by the river, meets your jump criteria. Read: listing emails arrive in Japanese and AI pulls out the details in English, marked unverified. Check: plain rules decide, listings land as Pass or Watch, and only the Jump sends a message. Filed: one card per place, next step book a visit for Saturday. The card photo opens into the house by the river, then a walk up the path to its gate, impression Promising. Screens are illustrated; places are fictional; scenes are AI-generated.

    Project 04 / 06 · Jul 2026 · Tracker

    Property Watcher reads every new property listing and tells me only about the ones that fit.

    No noiseSky messages me only when a place fits what I'm looking for

    The problem

    Listing alerts are noisy, and I kept losing track of places I'd already seen.

    What I built

    A tracker where AI reads each listing, clear rules score it, and every place gets its own file.

    Learned →Let AI read the messy stuff, but keep the final call in rules you can check.

    Python · SQLite · Claude · Sky · Grok Bot · Notion

  5. A 45-second fantasy-anime loop illustrating Agent OS. Three artisans build a bronze-and-cobalt phoenix: one finishes a feather panel, leaves a handover with the goal, what was done, what was checked and who goes next; another inspects the joining edge; the next maker seats the panel and releases it. Their completed creation awakens, unfolds its wings and takes flight, a fantasy metaphor for shared work. No blind handovers. Handover notes are illustrated with dummy data; scenes are AI-generated.

    Project 05 / 06 · Sep–Oct 2026 · AI teamwork

    Agent OS is the rulebook my AI agents follow when they hand work to each other.

    No blind handoversevery agent checks the last one's work before building on it

    The problem

    When several AIs work at once, the risk is one trusting another's unchecked work.

    What I built

    A shared rulebook and a handover note for every task: the goal, what was checked, and who's next.

    Learned →When AIs work as a team, the handover matters more than any single answer.

    Claude · Codex · Cursor · Sky · Sky Vault

  6. A 28-second loop told as a pixel-art RPG. A hooded adventurer stands at a guild hall notice board: where was I? In an archive of five glowing doorways full of scrolls: five AI apps, which one was it? The scene breaks into pixel tiles and becomes the Quest Board app: it reads every AI tool right on my Mac, groups the work into one quest per project, and one button, Open here, goes back in. The tiles burst into a workshop at dawn where the work was left: right where I left off. Quests are dummy; screens are illustrated; scenes are AI-generated.

    Project 06 / 06 · Oct 2026 · Mac app

    Quest Board answers “where was I?” across every AI tool I use.

    Five apps, one screenevery unfinished project in one place, with a button to jump back in

    The problem

    Getting back into a project meant digging through five AI apps for the right conversation.

    What I built

    A board on my Mac that turns every AI conversation into a quest I can pick up again.

    Learned →Plan it and write the tests first, and a team of AI agents can build a working app overnight.

    Python · SQLite · A local AI model · Claude Code · Codex

Live experiment · Paper trading only

Paper Desk. No real money.

My longest-running experiment, since April 2026. An AI agent runs a paper portfolio of large US stocks, picking them by a fixed set of rules and adjusting once a month. Two more agents are in design. The safety rules came first: an off switch, no code runs unless I have reviewed it, and an alert if the portfolio falls too far.

Paper trading only · No real money · Not investment advice · Delayed snapshot · date shown below

Loading the desk…

How I work

A real problem, a team of agents.

I pick a problem I actually have and build the fix with AI agents. They write the code, test it and check each other's work; I decide what ships. Each project leaves me with a tool I understand properly, and habits I carry into the next one. The AI tools I use most are Claude Code, Codex and Cursor.

Questions

Fair questions, short answers.

Is this a business?

No. It's a personal learning lab. I'm not offering services.

Are the projects real?

Yes. Each one is something I use. The screenshots and mockups use dummy data, and each caption says so.

Is Paper Desk investment advice?

No. It's an experiment that trades on paper only, with no real money. Nothing on this site is investment advice.

Follow along

What I learn, as I go.

I share what I built, what broke, and the tech worth knowing about.

Follow on XConnect on LinkedIn