Inspiration
Language courses usually decide what everyone learns next, then ask real life to wait. Probandolenguas reverses that relationship.
A learner has a changing universe of words, chunks, interests, errors and available attention. The useful question is not “Which lesson comes next?” but “Which new language gives this person the greatest increase in what they can understand and express now?” We turned that question into a curriculum function—and made the learning interface adapt to the learner’s actual context.
What it does
Probandolenguas is an app-agnostic adaptive learning engine. Spanish is its first full corpus and English powers the judge demo; the method is designed for many languages.
Its E(L) model treats a learner’s known language as a set L and ranks candidate words and chunks by marginal communicative reach: how many meaningful intents become expressible when an atom is added. Corpus frequency is one signal, not the curriculum. Personal interests, recurring difficulty, pragmatic value and fun-to-say material also shape the next pack. A protected human foundation keeps greetings, question words and other socially essential language from being optimized away.
The selected words become a complete learning loop:
- Meet them in contextual examples constrained to known language.
- Hear them in a generated two-person dialogue whose script is mechanically checked against the learner’s approved word universe.
- Retrieve them through varied inline exercises.
- Use them in text and voice conversation.
- Receive meaning-first feedback: preserve the conversation now, remember the error, and revisit it later through an evidence ledger and spaced exposure.
The same goal is compiled differently for the learner’s moment. Driving can be audio-only. A library session is silent. Low-energy or sick mode is short and passive without pretending exposure equals mastery. A phone gets vertical, social-style interactions; a computer gets longer writing; a TV can lead with video while the phone acts as controller. Learners can request a no-screen session or a printable page, and handwritten work can be photographed for AI-assisted review with confirmation before it changes the learner record.
Telegram is the first full channel adapter: chat, buttons, exercises, voice messages and generated audio stay in the conversation. The browser provides a signed, expiring workspace. WhatsApp and Instagram can reuse the same learning engine because transport is separated from pedagogy.
How we built it
GPT-5.6 was the coding and reasoning model used throughout our Codex workflow. Codex helped turn an evolving learning-method conversation into a working, tested system: a fail-closed rights pipeline, 5,500-intent CC0 corpus, staged optimizer, context compiler, channel-neutral event contract, Telegram voice adapter, generated exercises, signed links, durable learner state, scheduler, security controls, deployment automation and the public product experience.
The optional runtime coordinator uses the OpenAI Responses API for structured coaching. Voice notes use OpenAI transcription; the two-person listening path uses generated speech. The system separates the coordinator from deterministic policy: scope enforcement, rate limits, approved vocabulary, evidence rules and privacy boundaries do not depend on a model deciding to remember them.
Codex also acted as an engineering loop: implement, run adversarial and integration tests, inspect the real channel behavior, capture failures, and feed those failures into the next pass. The repository includes tests for the browser, optimizer, PostgreSQL learner state, Telegram text and voice, mocked OpenAI calls, scheduler behavior, rate limits and signed-workspace security.
Rights, safety and learner trust
The contest corpus contains only project-authored material dedicated under CC0. A source manifest, hashes and a fail-closed build gate prevent unknown or unapproved content from entering the corpus. The software is MIT licensed.
Learner identities are minimized and HMAC-indexed, links are signed and expire, model credentials remain server-side, and both per-learner and global cost limits fail closed. Passive exposure and productive evidence are stored separately; an exercise, transcription or handwriting interpretation never silently promotes itself into mastery.
Challenges and lessons
The central product correction was realizing that the engine—not a standalone screen—is the product. Once pedagogy, learner state and safety became shared services, Telegram, the browser, future messaging channels, audio-only sessions and printable work became different expressions of one learning relationship.
The optimizer also taught us that pure mathematical search can remove socially essential words. The resulting staged design is stronger: humans protect the communicative foundation; the model audits it and optimizes the expanding edge. That is the method in miniature—AI expands human judgment instead of replacing it.
What's next
Next we will activate production WhatsApp and Instagram adapters, add richer cross-channel identity export and deletion, grow the reviewed contributor corpus, run the preregistered semantic evaluation, and bring the same method to a third language. The long-term product is a learning relationship that follows the learner—across languages, devices, energy levels and real life.
Built With
- codex
- docker
- gpt-5.6
- next.js
- openai
- postgresql
- python
- react
- responses-api
- telegram-bot-api
- typescript