The 14 projects I run as the person in charge.
Inspired by Pieter Levels (@levelsio), I'm sharing the 12-month process on X.
All as CEO. Cited from @levelsio: "I do hire contractors for specific things."
A project of Athena.
🔗 Coming soon
A service that automatically turns performance data (MIDI) into clean, readable sheet-music PDFs.
Copyright and royalty management for songwriters and composers.
🔗 Coming soon
SaaS for running concept cafés: cast sales aggregation, regular-customer CRM, shifts, and payroll. In development at a separate company.
🔗 Coming soon
JPYC-integrated; a 3% fee returns 97% to artists. Built on the "Tamago Index" (a logarithmic visualization of social-contribution value). A MUGI business.
🔗 Coming soon
AI profile photos for job hunting and career changes. Fine-tuned on Japanese faces + industry-specific dress codes. An industry-specialized take on @levelsio's PhotoAI.
🔗 Coming soon
A series on AI × hardware — the exact setup I actually use.
A job board specialized for AI / indie hackers. A Japan-market version of @levelsio's RemoteOK.
🔗 Coming soon
A complete record of the 12-month process. A strategy book in the same vein as @levelsio's MAKE Book. Planned for May 2027.
🔗 Coming soon
A guide and directory for Japanese concept cafés — profiles for ~1,000 shops across Japan plus visit reports.
R&D combining specialized small AI models to reach frontier-level performance in a focused domain — run locally at low cost, keeping sensitive data in-house. First domain: music copyright & royalty data analysis.
🔗 Coming soon
A self-built sensor that uses an AI camera to count street foot traffic 24/7, turning it into people-flow data — an explanatory variable for a business-conditions index. Built from scratch, from camera to AI, without off-the-shelf products.
🔗 Coming soon
A bilingual media site introducing Japanese anime soundtrack ("gekiban") music to the world — feature articles on works and composers, plus soundtrack event listings at home and abroad. Run by Athena.
A service that lets individuals book ads on large street-side digital screens entirely online — upload a video, pick a time slot, done. Play counts are guaranteed, with refunds if they fall short.
🔗 Coming soon
A damages claim over a former employee's unlawful conduct. Pursuing multiple cases in parallel, representing myself (pro se).
CueFlow — For our fourth evaluation piece, an outside expert has started reviewing the quality of a string (cello) part. After wind instruments and piano, we are now checking the conversion on strings too, to confirm the output holds up to the same standard as the instrument changes.
CueFlow — Based on an outside expert's answers about tuplet notation, we improved how triplet beams are split into groups of three and how the note value is chosen for tuplets with many notes, and shipped it to production. Each fix went through repeated AI code reviews before being merged into one release, and the updated scores are now back with the expert for re-checking.
CueFlow — Notation standardization completed a full loop: ask an outside professional, fold the answers into production, then re-verify in the same screen. Decisions we had been handling piece by piece — how rests are written, how tempo is shown — are now fixed as product-wide standards, and the reviewer re-runs the conversion in their own account to confirm the result. The review itself is now a fixed form too: the number of spots that must be fixed before the score is usable, the reviewer’s own estimated correction time to a deliverable state, and a verdict of usable / usable after fixes / not usable. We also extended the review to a piece with different instrumentation, with part-by-part checks continuing. For that piece the reviewer went through every page, and we corrected how notes are grouped, how rests are written and which way stems point according to the time signature. The basic theory rules now apply across every piece rather than one, and we are widening their coverage time signature by time signature. We also fixed the review screen showing nothing but a run of numbers on pieces with long stretches of rest: long rests are now shown as a single block, and a comment can be attached to the start of that range. A third piece — a solo piano work — has now been through the same loop, and of the 100-plus comments received across two pieces we corrected and shipped roughly 60 percent. Fixes covered note grouping, how rests are written and stem direction by time signature, plus very short notes that were coming out malformed; what remains — writing that depends on reading the key correctly, and unusual tuplets — is queued for the next round. We also agreed to leave out of review the things MIDI alone cannot decide, such as repeat instructions and which hand plays a passage, which sharpens the comments we do get. A full logo and icon set for the service is now in place. Next, we took on the item we had queued — choosing accidentals that fit the key — and merged it into the product together with correct grouping for tuplets that start with a rest and hiding the filler rests that appear when voices are cleaned up. On the second piece, 54 of the 57 checks derived from the reviewer’s comments now pass (the remaining 3 are complex tuplets), with all of the roughly 2,200 automated tests still green. Read more →
Reward Distribution System — Rolled out a fully revised beta spec to production, with a release procedure that backs up existing data, reconciles it before and after migration, and passes every automated test and CI run before going live. We also built a verification module that credits stablecoin deposits only once they are confirmed, and only once (test network only; commercial use not yet decided).
Concafe OS — Ran a design and code review focused on where the time-clock terminal meets the core system, surfacing 13 issues to fix before go-live, such as clocking in on consecutive days and recovering from network outages. Eleven were resolved through an automated AI fix loop; the rest are parked pending business decisions and expert advice.
Copyright Management System — Built a "disclosure portal" into the existing system so rights holders can check contract-based revenue and documents at any time. It is designed so the operator cannot hide information, backed by a tamper-evident audit log. We also added a hiring aid that lets applicants from a sign-up form be auditioned and rated on the spot. That hiring aid now covers first-round interviews too: slots are pulled in automatically from the booking calendar, impressions and decisions are recorded, and the people involved are notified. Interview recordings are kept in private storage only authorized staff can open, with every playback and deletion logged. We also added inbound inquiry handling: messages addressed to the rights holder are pulled into the system, foreign-language ones get an automatic Japanese translation and a summary of the key points, a draft reply is prepared on the spot, and the reply can be sent straight from the system.
Postari — To avoid repeating the configuration gaps found during production testing, we added a production-readiness checklist and an automated external check to run before every release. Next up is Instagram integration, with implementation and app-review preparation (screencast and data-handling policy) under way.
Postari — Completed an end-to-end billing test in production, covering both the monthly plan and pay-as-you-go usage charges. We confirmed that everything from sign-up with a real card through posting to invoicing works correctly, and set up the essentials for production operation, including migrating the image-delivery domain and adding qualified-invoice (tax) support.
Lakho — Designed and began building a "review dashboard" that lets a person check the recorded footage and correct the counts. Reviews are logged with full history as a single source of truth for corrections, and will later serve as training data for AI to make those corrections automatically. We also added a system that flags gaps and quality issues in the observations, so time spans we could not measure are never mistaken for a genuine "zero."
Konnavi (こんなび) — Coverage of the greater Tokyo area is now essentially complete. From here we pick up newly opened shops nationwide each day, extending our listings out to regional cities. Read more →
Lakho — Assembled a second measurement unit and finished preparing it for deployment in another city. We also built a server and dashboard that gather readings from multiple sites into one place for aggregation and visualization, plus step-by-step guides covering assembly, camera setup and on-site installation so anyone can bring a new unit online the same way.
Concafe OS — Finalized the overall design that unifies attendance, point cards, and register data — long handled by separate systems — into a single platform, and moved into implementation. To avoid disrupting the shops, the old and new systems will run side by side and their totals reconciled before we switch over.
Lakho — Our self-made people-flow sensor has now logged about 10 days of real-world foot-traffic data, and we ran our first full statistical analysis. By checking the counts against actual footage we verified accuracy, and began designing a two-stage system that corrects for over-counting.
Reward Distribution System — Completed an external engineering design review and shipped every recommended safeguard to production (protection against invalid balance operations, an append-only audit trail of all operations, and more). Also advancing the design for external payouts.
Gekiban Portal "GBmedia" — Running a bilingual media site that introduces Japanese anime soundtrack music to the world, with regular feature articles on works and composers plus soundtrack event listings at home and abroad. Read more →
Copyright Management System — Built a pipeline that automatically ingests royalty statements from 20 publishers and verifies the amounts against the real data. Now researching collection societies worldwide, designing features to recover overseas royalties without leaving any uncollected, and beginning to ingest and analyze real overseas distribution (streaming) data.
World-class Homegrown AI — Kicked off a new project that combines specialized small AI models. First domain: music copyright & royalty data analysis.
CueFlow — Fixed, at the root, a bug where pieces could be cut off partway through — so scores now run all the way to the final note. Also began building a system that logs and reviews conversion quality to keep improving it (review dashboard). Read more →
Konnavi (こんなび) — Officially launched the fast, bilingual (JA/EN) site at concafe.help, now expanded to ~1,000 café profiles across Japan plus our visit reports. Read more →
BlackPrince Lawsuit — Filed a full rebuttal to the defendant's briefs. Appealed the dismissal in the lost-earnings case and submitted briefs in the appeal and the main case. Read more →
Pieter Levels (@levelsio) — Serial entrepreneur, levels.io/projects/
Progress, failures, and numbers — I share it all on 𝕏.
In 12 months (planned for May 2027), I'll tell the whole story in a retrospective note.