Current

Inspiration

Your field changed this morning. Your course did not. AI has given us effectively infinite explanations, but explanations are not the same as understanding. Chat interfaces are great at removing friction, including the productive friction that makes learning stick. Traditional courses provide structure, but they quickly become outdated in fast-moving fields. I built Current to sit between those two extremes: a source-grounded learning system that adapts to the learner, remembers what they misunderstand, and updates its curriculum as the underlying material changes.

What it does

Current turns documentation, papers, links, and uploaded files into structured learning paths: Sources → Read → Recall → Apply → Reflect → Review Learners must reconstruct ideas without looking at the source, apply them to concrete decisions, and reflect on where the knowledge belongs. Current evaluates understanding rather than exact wording, records concept-level misconceptions, adapts the next activity, and schedules future retrieval practice. A research agent monitors the sources behind each path. When information changes, Current shows the stored and latest evidence, identifies affected concepts, and proposes a curriculum update for approval instead of silently rewriting the lesson. Current Classroom extends the same system to educators. Teachers can see shared misconceptions, assign targeted reviews, personalize examples around student interests, and receive evidence when students demonstrate understanding. The context can change from basketball to spaceflight; the objective, source, and scoring rubric remain consistent.

How we built it

Current is built with Next.js 16, React 19, TypeScript, React Flow, and the OpenAI Responses API. It is deployed through OpenAI Sites. I treated GPT-5.6 as a system rather than using one model for everything: GPT-5.6 Sol plans learning paths and evaluates source changes. GPT-5.6 Terra authors complete lessons from approved sources. GPT-5.6 Luna provides fast, structured recall and application feedback. Strict structured outputs keep model responses compatible with the product runtime. The local-first learning runtime stores progress, notes, uploaded sources, reflections, review schedules, support preferences, and misconception memory. Deterministic fallbacks keep the demo reproducible when live model access is unavailable. Codex was part of the entire development loop: challenging the original product thesis, implementing the application, testing user flows, debugging hosted behavior, and repeatedly removing dead controls and generic AI patterns.

Challenges we ran into

The hardest challenge was product direction. My first attempt looked like another generic AI dashboard. I scrapped it and rebuilt Current around one concrete claim: progress should mean demonstrating an idea, not scrolling past it. The second challenge was state. The learner loop, review queue, research updates, learning map, and Classroom all needed to reflect the same evidence without feeling like disconnected demos. The third challenge was trust. Research agents cannot be allowed to silently change educational content, especially in a classroom. Current therefore exposes source evidence, model provenance, affected concepts, and teacher approval controls. Finally, I had to make live AI behavior reliable enough for a judged demo. That led to strict schemas, task-specific model routing, visible provenance, deterministic fallbacks, a complete reset flow, and automated coverage across the critical workflows.

Accomplishments that we're proud of

Current is not a mockup or chat wrapper. The complete workflow functions end to end: Source-grounded path generation Generated Read, Recall, Apply, and Reflect lessons Concept-level misconception memory Adaptive support and spaced review Autonomous source monitoring with human approval Classroom assignments and targeted interventions Student evidence returning to the teacher in real time Thirty passing product and API tests

What we learned

I learned that educational AI becomes more useful when it stops trying to replace effort and starts deciding where effort is valuable. I also learned that personalization should respond to demonstrated needs, not assign people fixed “learning styles.” Current adapts support formats based on task fit, learner choice, and previous outcomes while preserving shared objectives. Most importantly, I learned that the strongest educator workflow is not “AI teaches the class.” It is AI makes learning evidence legible enough for a teacher to act on it.

What's next for Current

The next steps are learner profiles, durable cloud synchronization, teacher-authored curriculum constraints, optional focus tools, and classroom pilots with educators. The larger goal is to help students and lifelong learners pursue their own interests without sacrificing rigor, evidence, or durable understanding.

Built With

• codex
• gpt-5.6
• next.js
• node.js
• openai
• react
• typescript