LearnCraft Spanish Quizzing App

LearnCraft Spanish is a premium Spanish coaching business focused on genuine fluency for adult learners. It has always been unapologetically grammar-focused, targeting motivated English speakers who want to do the hard work it takes to be able to connect meaningfully with the native Spanish speakers in their daily lives.

Since the company was established in 2014, I had helped out part-time with early messaging, brand differentiation, and design work. I was brought on full-time in 2017 as head of business so that Timothy could focus on creative and pedagogical improvements. Over the next four years, I built out the record-keeping systems, established SOPs and quality control measures, hired and onboarded staff, organized the leadership and middle management layers, and increasingly, leaned on proprietary no-code software solutions to handle the growing volume of data and workflows.

The company's ideas kept outrunning its tools. Timothy had long envisioned a quizzing tool that would allow us to save a bank of sentence examples in English and Spanish, tagged by their specific vocabulary and grammatical concepts. Originally the intent was to assist our native-speaking coaches in their one-on-one sessions, but we quickly realized that students would benefit as well if they could use it directly.

We were able to use our existing no-code data platform to prototype a structure for tagging sentence examples, but none of our tools could filter, collect, or display them in a way that would be useful for our coaches or students.

Around 2021, I started teaching myself how to code, both to support, extend, and integrate our existing tools and to develop the capacity to build any new ones we would need. One of our students helped rig up a simple React app that could pull the data from the API and filter the examples with JavaScript. It was a single component, but already enough to be useful, and from there we were hooked.

I took the proof of concept and began building it out. We created a full stack with a strong separation between frontend and backend, used Auth0 for authentication with role-based access control, and deployed it to a custom domain. Eventually, our students were able to search our examples by specific vocabulary words, add them to a personal collection, and review them with a spaced repetition system that would increase or decrease the review interval based on their self-reported difficulty rating. We also added curated sets for each published lesson, and soon created audio-first reviews as well.

(The video below is actually from February of 2025, so it shows more functionality than we had at this point.)

With every new feature we added, the cracks in our system showed more and more clearly. It turns out that building a public-facing product on top of a rate-limited API that can only join tables via lookup and summary fields isn't a good long-term data strategy. Our data platform was able to express the relationships we needed, but it wasn't designed to query across multiple tables at once. My workaround solution had been to pull large data sets and cross-reference records by ID directly in the browser.

I also noticed more and more bugs and repetition creeping into the codebase. Originally I had built very quickly, using plain JavaScript and React, and most of the logic was written directly inside the monolithic components. This meant a lot of duplication, more places to check when a new feature was added, and more surface area for bugs to creep in.

In the meantime, student records, which were increasingly being reviewed and updated through the admin side of the same app, sat in a completely separate database from the vocabulary and example data, so integrating the student's quizzing data with their coaching sessions was completely impossible. As the user base grew, we became increasingly concerned about the vendor's uptime and the API rate limit. We knew we needed to migrate to a proper database solution.

The deeper challenge was that I had been building everything more tightly coupled to the existing data platform than I realized, and even if we could have migrated the data, the queries scattered throughout the codebase, the proprietary syntax, and other vendor-specific details would have prevented us from using it.

I had to either accept the limitations and compounding technical debt, rewrite the entire app from scratch (freezing all features until the whole migration was completed), or find a way to change the code and the data dependencies incrementally under live traffic.

It was clear that the path ahead of us would require more than just my full-time effort. Even if we could find a way to mature the codebase and migrate the data without blocking features, there was no way I would have time to pursue both new feature development and migration.

We started recruiting for a full-time engineer to help me with the work, and in late 2024, we were able to find a great fit in Blake Willis, a graduate of the Atlas School program.

Adjusting the workflows and coding practices for collaboration was another exciting challenge for me, since I had been working solo for so long. We quickly realized that we needed clearer quality control and monitoring practices to hold our work to a high standard.

All of these factors converged in some very deep and thoughtful technical discussions between the two of us and the company leadership, and we ended up with all of the quality and observability metrics we needed, plus a new strategic plan.

I had started rolling out TypeScript in early 2024, the first move in what became a much longer plan. My intention was to complete the rearchitecture-in-place with the lights on the full time, and to get it done the right way so that it would never have to be done again.

Early on, the process started in one direction and then had to double back when the result either wasn't rigorous enough for long-term maintainability, or ended up adding more indirection than value.

In the end, we accomplished the following:

As the core chunks of business logic were isolated and extracted, we found that we could recombine some of our existing logic to unlock specific functionalities that had been intended since early in the project, but were previously impossible due to both the data handling we had formerly implemented and the poor codebase isolation.

After reworking the core filtering and quizzing logic, we were able to implement a completely customized quizzing interface that allows students to quiz any arbitrary vocabulary words, parts of speech, verb conjugations, and more, with the certainty that all the vocabulary will be fully within what they have already learned. Due to the shared domain logic, this could be done in both flashcard and audio quizzing modes.

Data handling had been moved entirely to the backend, and queries could be shuffled with a seed for server-side pagination while preserving the same randomized order. A server-side cache was also implemented to reduce the rate-limit impact while the data was preparing for migration.

At the end of 2025, I determined that I was ready to move on from LearnCraft Spanish. A new opportunity had presented itself, and I was ready to take the next step in my career.

My final months were spent preparing the codebase for long-term continuation by Blake and any future engineers, as well as for the increasing impact of AI tools on the codebase. One positive outcome of the restructuring was that the codebase patterns and practices were clear enough that AI tools imitating their surroundings would naturally follow our best practices, and we were able to leverage them more and more effectively through 2025 and beyond.

To prevent drift or further regressions, I completed a thorough documentation of the specific architectural decisions and testing patterns that we had developed, as well as the full PR review process and completion checklist. Boundaries and dependency directions were documented in clearer detail and brought up to current practice, and the linter rules were updated to enforce them. I also wrote a CI workflow for Claude Code to run a PR audit against a specific checklist and hard block merges on failing reviews.

The frontend repository has always been kept public, so feel free to see for yourself:

The slow-moving migrations have finished their course in the time since I left, and the data is finally safe in an owned PostgreSQL database. The team continues to ship new features and improve the product, and public-facing functionality is completely in the new architecture.

The app has also significantly increased the lifetime value of the LearnCraft Spanish customers. In a survey of 205 users, 78% said they would be “very disappointed” to lose the product (the traditional Sean Ellis PMF benchmark is 40%), and it is regularly mentioned directly in student reviews and surveys.

While most of the functionality is available only through the coaching program, the quizzes for the public courses are available to use for free: