Product · 2026.02 – present

Lingxian — online jianzipu editor for the guqin

A notation editor and small SaaS for guqin players, built solo from February 2026 and now in invite-only beta: 94 registered users, 160 scores, 144 PDF exports and 20.7% week-2 retention (as of September 2026).

Role
Founder — product, engineering (AI-assisted) and operations, solo
Period
2026.02 – present
Stack
Next.js 14 · React 18 · TypeScript · Prisma + PostgreSQL · VexFlow · Auth.js · Qwen (LLM, intent only) · WebSocket + Yjs · pnpm monorepo
Scale
94 users · 160 scores · 144 PDF exports (beta, 2026-09) · 20.7% week-2 retention (n=82) · 1,200+ commits · ZH + EN site, live at lingxian.app
Links

Problem

Guqin players shared scores by hand-copying and phone photos. Tablature (jianzipu) tells you which string, which position and which finger — but not rhythm, tempo or weight. Players decide those phrase by phrase, and the decisions live only in their memory. Three things go wrong:

  1. Decisions are lost. Where to break a phrase, where to slow down: you settle it line by line, but the notation can’t record it. Months later you reopen the score and can’t remember why you chose what you chose.
  2. Versions are opaque. A piece goes through several revisions, each saved as a new file. Finding which version changed which note means laying two printouts side by side.
  3. Feedback is messy. You send a score to your teacher; they screenshot it, circle things in red, send it back. Three rounds later nobody can say which version or symbol is being discussed.

I know these pains first-hand from four years as a performance major. No existing tool treated jianzipu as structured data, so I defined, built (with AI coding tools, Next.js + PostgreSQL) and launched one from scratch.

Decisions

Model the score as data, not as a drawing. Every note is a Token (string, hui, fen, technique, duration, ornaments); sections and documents compose them. This single decision is what makes staff rendering, version diffing, collaboration, PDF export and — later — automatic transcription possible from one source of truth.

Let sign-up applications write the roadmap. I coded 119 sign-up applications into 10 need categories. 34 asked for the same thing: convert numbered notation (jianpu) into guqin notation automatically. I framed it as candidate generation + ranking: a rule engine enumerates every valid fingering, beam search picks the best-scoring sequence, and a candidate panel lets the player override. Every override is logged — ~3,800 edit events so far — as training data for a future ranking model.

Measure the funnel one step at a time. Sign-up → create → save → usable score → export, each counted separately. The largest loss was creators who never produced a usable score: 53% (30 of 57, mid-September). It matched the most common feedback, “input is too hard”, so I shipped keyboard-style and batch input. The step now loses 35% (24 of 69) — no control group, so I report it as a before/after, not a causal effect.

Chase the small leaks too. Only 19 of 35 approved applicants (54%) redeemed their invites. The cause was a 2-day token expiry plus invites landing in spam; longer expiry and one-click resend fixed it.

Price before you can charge. I designed a four-tier plan structure and opened two tiers online, but paid billing is on hold until there is a Chinese business entity. To validate demand anyway I gifted plans: 2 sponsorships and 5 purchase-intent sign-ups on the pricing page.

An AI assistant that cannot touch the score on its own (Jul 2026 →). The in-editor assistant started as an answer-only chatbot. I rebuilt it as a natural-language operation entry point where the LLM (Qwen) only parses intent and writes replies; deterministic tools do every calculation and edit, behind a check → preview → confirm → execute → undo state machine with stale-snapshot detection. I built a 74-case Chinese intent eval set (ambiguous, colloquial and high-risk negative cases) and iterated prompts v1.0 → v1.4 by error type: 96% intent accuracy (71/74), 100% valid JSON, zero high-risk false triggers. All 38 live interactions from 13 users succeeded — but only ~13% of users tried it, so the next step is adoption, not accuracy: adoption / undo / trigger-rate metrics are defined and the entry point is moving into the toolbar.

Result

As of September 2026, seven months in: 94 registered users, 160 scores, 144 PDF exports, 20.7% week-2 retention (n=82) and 19.7% month-2 retention (n=76), on an invite-only beta at lingxian.app. The product renders tablature, staff and jianpu together as you type, converts jianpu to jianzipu with per-note alternatives, diffs versions side by side, supports real-time co-editing with comments, assembles textbooks for print, and ships a browser tuner, a 149-entry fingering reference and a 64-article blog for SEO.

Engineering footprint: 1,200+ commits since February 2026 (see the activity grid), a pnpm/turbo monorepo with web, api and mobile apps plus a shared music-theory package, Prisma migrations, Vitest, Docker images on GHCR behind nginx. The code is public at github.com/Zhuohengli03/lingxian, a mirror of the private development repository whose history the activity grid shows.

← Back to home