A personal learning app for the small moments between everything else.
Interlude turns a spare few minutes into a focused lesson: read or listen, check what you remember, save a thought, and pick up where you left off. It is a full-stack progressive web app designed for an iPhone and a laptop, with one private owner per deployment.
Built by Kai Takeuchi, with AI-assisted development and editorial work through Codex.
I kept reaching for my phone in the gaps in my day, and five spare minutes often turned into doomscrolling. I wanted to spend more of that time learning about things I care about: AI, software, healthcare, history, economics, communication, and being a better parent and partner.
I'd tried learning apps before, but the subscriptions felt too expensive for what I wanted. YouTube and podcasts helped, but I still had to choose something, switch apps, and remember where I left off. I wanted one place to pick up a useful lesson, listen while doing dishes or walking, and read or save an idea on my laptop.
I also wanted to get better at building applications. Making my own learning app let me do both: learn about my interests and learn by building something I'd use. So I figured, why not build it myself?
Interlude is that ongoing experiment. It is designed around short lessons, a small running budget, and a clear place to stop. There is no infinite feed or streak penalty. The aim is to remember and use what I learn, not just finish more content.
The current library, captured October 9, 2026 from an isolated local preview of the private app. It uses a Demo account with no personal learning history. The three companion courses shown here are newer than this public source snapshot.
The gallery below was captured September 15, 2026. Its profile, progress, calendar, and notes are synthetic examples. The player uses a silent local audio fixture to show its controls; these images do not demonstrate voice quality. Click an image to view it at full size.
| Pick up your current lesson | Read or listen at your own pace |
|---|---|
 |
 |
| Follow a few learning paths | See your learning rhythm |
|---|---|
 |
 |
- Start quickly: Today surfaces an in-progress lesson or the next lesson in an active course.
- Read or listen: Original cited lessons, saved narration, adjustable speed starting at 1.5×, seeking, and a persistent player. Tapping the footer player opens its lesson.
- Keep sessions bounded: Approximately five-minute lessons at normal speed and a deliberately selected queue of up to three lessons.
- Practice optionally: Recall questions, reflections, practical challenges, follow-up questions, and short spoken input with an editable transcript.
- Follow your curiosity: Search, pause and reorder courses, save topics and sources, and prepare standalone lessons from reviewed evidence.
- See progress: Durable cross-device positions, a contribution-style activity calendar, and separate completion, recall, and application records.
- Keep evidence and history: Dated citations, retained source provenance, immutable lesson versions, and exportable learning data.
- Control costs: Reserve usage before provider calls, reuse saved audio, retain uncertain attempts, and enforce application spending and storage limits.
The private app now has six courses and 36 lesson texts. This public source snapshot includes the original three courses and 24 lessons below.
| Course | Focus | Lessons |
|---|---|---|
| AI for product managers | Problem selection, inference, retrieval, uncertainty, evaluation, adoption, and a product brief | 8 |
| AI for software development | Client/server/data, APIs, assistant context, code review, tests, deployment, and a feature plan | 8 |
| AI in pharma & healthcare | Workflow mapping, commercial insights, field support, content review, patient access, evidence, and a pilot | 8 |
The live app also has three original companion courses inspired by concepts in Chip Huyen's AI Engineering, added October 9, 2026:
| Companion course | Focus | Lessons |
|---|---|---|
| Model decisions | System design, tokens and context, model selection, and prompt specifications | 4 |
| Evaluate and improve | Evaluation sets, AI judges, choosing an intervention, and dataset quality | 4 |
| Reliable systems | RAG diagnosis, agent boundaries, latency, and production feedback | 4 |
Each companion lesson pairs short reading with an original diagram, a narration script, and optional practice. These are original lessons grounded in cited primary sources, not copied book chapters or illustrations. Their source and diagram renderer are not yet part of this public snapshot. New narration is pending recovery of an interrupted audio request; publishing text does not mean its audio is ready.
Generated audio belongs to an individual deployment and is not included in the repository. Healthcare examples are fictional teaching scenarios; they are not patient-specific advice or claims of deployed clinical outcomes.
New lessons follow the lesson voice guide: teach concepts directly, keep routine publisher attribution in source notes, and name sources aloud only when their identity matters. Existing released texts and recordings are preserved until an explicit revision is published.
| Layer | Implementation | Why it fits this project |
|---|---|---|
| Interface | TypeScript, React 19, Next.js App Router conventions | Shared types and reusable components across phone and laptop |
| Build/runtime adapter | Vite 8 + vinext, Cloudflare Vite plugin | Runs the Next-style application on a Workers-compatible runtime; retains the existing Sites integration |
| UI | Tailwind CSS 4, Radix/shadcn-style components, Lucide | Consistent responsive controls without building every primitive from scratch |
| Server | Cloudflare Workers through Sites | A small HTTP application with managed deployment; no separate always-on application server |
| Data | D1 SQLite, Drizzle schema and SQL migrations | Durable relational state with atomic updates and an inspectable schema |
| Audio and source artifacts | Private R2 object storage | Large files stay out of the relational database; audio is served through authorized routes |
| Narration | Google Cloud Text-to-Speech, Chirp 3 HD | Saved natural-sounding narration with separate conservative character accounting |
| Optional interactive AI | OpenAI Responses API and transcription | Bounded questions, research, lesson preparation, and short voice drafts |
| Content maintenance | Codex task + trusted publication artifacts | Lets the personal deployment use plan-based research and authoring, while keeping optional runtime API work separate |
| Checks | TypeScript, ESLint, Node test runner, SQLite fixtures, workerd | Tests state changes, evidence checks, recovery, and the actual Worker audio adapter |
This is a Next-style app built with vinext, not a standard next build deployment. Production authentication depends on the trusted Sites dispatcher. Another host needs a real authentication integration before it can safely run this application.
Read Architecture and decisions for the tradeoffs, boundaries, and alternatives considered.
This grew through small implementation trials and feedback from personal use. These were development experiments, not controlled user studies. The most useful lessons came from trying the complete workflow, including what happens when something fails.
| What we explored | What happened and what we learned |
|---|---|
| A two-lesson pilot before a full library | Started with two lessons to exercise sign-in, saved positions, audio, and phone use. Expanded to eight, then 24 lesson texts after playback checks and owner feedback. Prove the learning loop before scaling the catalog. |
| A home screen that answers “what next?” | Personal use exposed the friction of finding the current lesson. Today now chooses a saved unfinished lesson or the next available lesson. A separate spacing fix gave the home cards consistent gaps. The entry point deserves as much attention as the library. |
| Faster audio and an expandable footer player | Changed the default to 1.5× and made the footer open the correct lesson while preserving the media element and queue. The owner confirmed both refinements worked. A small interaction can matter every session. |
| OpenAI narration, then Google Chirp 3 HD | The first course used saved OpenAI narration. We played Charon and Kore samples, tried a full Charon lesson, and selected Charon for new narration. Existing files remain playable. Voice choice and cost accounting are separate decisions; saved replays avoid new synthesis. |
| Testing in the actual Worker runtime | Mocked provider tests missed a fetch redirect-mode incompatibility that failed during Google authorization. Reproducing it in workerd led to a fix and a runtime-specific test. Passing unit tests did not establish platform compatibility. |
| Recovery after an interrupted narration | A live interruption exposed an unreadable error and a missing lesson-level recovery action. Both were fixed, but replacing a genuinely missing Chirp segment is still unfinished. A lost response does not prove the provider did no work, so an automatic retry can be the wrong recovery. |
| Plan-based authoring alongside API features | The personal weekly workflow uses a Codex task and reviewed publication artifacts; optional in-app AI uses API credentials. This reduces the need for runtime generation but depends on an awake laptop. A refresh button alone was not an unattended schedule. |
| Learning measures beyond completion | Completion, recall, application notes, and active time have separate records. A contribution-style calendar replaced the idea of streak pressure. These are useful signals to inspect, not proof of improved retention or less scrolling. |
The recurring lessons: make the next action obvious, preserve the user's place, keep provider work durable, and make evidence and uncertainty visible. Long-term learning outcomes and total monthly costs still need measurement.
The build retrospective records the trials, observed results, decisions, and remaining questions, including lesson versioning, embedded media, and preparing a reproducible public repository.
Requirements: Node.js 24 and npm. The versioned lockfile is included.
git clone https://github.com/ktakeuchi21/interlude.git
cd interlude
npm run install:ci
npm run build
npm run db:migrate:local
npm run devOpen the Local URL printed by the dev server, normally http://localhost:5173. Local sign-in uses a clearly separated development identity; it does not sign into the private production account. Reading, progress, and exercises can be explored without provider credentials. Narration and optional AI actions require your own server-side configuration.
See Setup and deployment for migration behavior, Google project configuration, optional secrets, and production authentication requirements.
The personal design target is less than $10 per month. That is a target, not a guaranteed bill.
The implementation caps OpenAI API commitments at $8 per UTC month, Chirp at 900,000 conservative UTF-8 bytes over a rolling 32-day window, and audio storage reservations at 2 GB. These are application controls, not account-wide provider billing limits. Replaying an existing audio file does not synthesize it again. Hosting, storage, other account usage, taxes, and provider pricing still matter.
Original course authoring and the personal weekly refresh can run through a Codex task using the owner's plan. The configured personal workflow runs Mondays at 8 AM Mountain and requires an awake laptop with Codex running. Cloning this repository does not install that automation. The public publication list starts empty, and the scheduling setup is documented in Weekly refresh.
In-app AI questions, research, lesson generation, and transcription use configured API credentials and accounting. They do not automatically consume ChatGPT subscription credits. See Operations and cost controls.
This is a working personal project under active development, published as a source reference rather than a multi-user service.
- The three foundation courses and core learning loop are implemented.
- Phone playback and cross-device use have been exercised in the personal deployment. Device behavior still needs checking on each target iOS/browser combination.
- Live generated audio is not bundled; a new deployment starts without it.
- Interrupted Chirp work can remain held if a provider result is missing. Saved-only recovery is available; automatic retries and a complete Chirp replacement workflow are intentionally absent.
- More live quality evaluation is needed for researched answers, generated lessons, transcription, and long-term cost assumptions.
- Offline downloads, Audible account integration, and multi-user accounts are outside the current implementation.
Roadmap · Verification · iPhone playback checklist
npm run typecheck
npm run lint
npm test
npm run test:worker
npm run check:public
npm run buildThe core suite currently contains 127 tests. Its providers and keys are synthetic; it does not send paid requests. GitHub Actions runs the checks and build without production secrets. See Contributing before changing state, accounting, or authentication.
app/ App Router pages, private API routes, authentication adapter
components/ Reader, player, library, exercises, settings, UI primitives
lib/ Lessons, evidence, progress, provider and accounting logic
db/ + drizzle/ Schema and ordered migrations
public/ PWA manifest, icons, online-only service worker
scripts/ Local setup, tests, publication preparation, public-source check
tests/ Core fixtures and loopback-only HTTP tests
docs/ Product, architecture, experiments, screenshots, setup, operations
This repository starts from a sanitized source snapshot. Production identifiers, owner-specific weekly publications, private operational logs, databases, credentials, and saved media are excluded. Public source does not grant access to the personal app.
Third-party notices are retained in THIRD_PARTY_NOTICES.md. No general reuse license is granted for Interlude's original code or lesson text at this time. The public repository is available to inspect; third-party components retain their own licenses.