Vidbyte
accelerate human learning with ai
I started building Vidbyte after noticing a problem with most learning products: they make it very easy to consume information, but much harder to know whether that information became something you can actually use.
AI makes this problem more subtle. It can explain almost anything clearly, summarize a book in seconds, generate a study plan, and answer the question you were stuck on. The experience feels productive because the output is polished. But a clear explanation is not the same thing as durable understanding. It is possible to leave an AI session feeling like you learned something and then discover later that you cannot recall it, explain it, or apply it without opening the conversation again.
The problem is not that we lack access to information. It is that we are still bad at converting exposure into mastery.
Vidbyte is an AI-native learning platform built around that conversion. The idea is to optimize for learning velocity: not how quickly someone finishes a video or moves through a course, but how quickly they build durable, transferable understanding that holds up when the notes are gone and the problem changes.
Vidbyte is an attempt to connect those pieces into one system. The products we are building include learning pathways, diagnostic conversations, active practice and assessment, research tools, and project-based synthesis. Each component should be useful on its own, but the larger goal is for them to share a model of what the learner is trying to do, what they already know, where they are struggling, and what kind of practice would move them forward.
The design principle is that every feature should respond to a specific failure mode in learning. If familiarity is being mistaken for mastery, the product should ask the learner to retrieve or apply the idea. If a goal is too vague, it should help turn it into a sequence of concrete outcomes. If someone can solve a problem only in the exact form in which they studied it, the system should create opportunities for transfer.
This philosophy also shaped the developer work around Vidbyte. While building the learning system, I kept running into the same problems when building AI agents: long conversations become noisy, agents lose progress, retries are expensive, and a successful-looking response does not prove that the underlying task was completed correctly.
That led to the Vidbyte SDK, an open-source Python framework for building, evaluating, instrumenting, and distributing AI workflows. It includes composable agents, tools, deterministic middleware, context management, durable sessions, checkpoints, traces, evals, pipelines, workflows, prompts, and MCP integration. The goal is to make an agent workflow more like an inspectable system with budgets, state, permissions, and evidence, rather than a single model call hidden behind a polished interface.
It also led to Vidbyte Skills: portable skills that bring learning and reasoning interventions directly into coding-agent sessions. A skill can notice when the assistant is making too many decisions, ask the user to retrieve what they just built, require an explanation of why plausible alternatives fail, or turn the important ideas from a session into a short practice exercise.
The common thread is that AI should increase a person's capability, not quietly replace the parts of the process that create capability. The best experience will not always be the one with the least friction. Sometimes a useful product has to ask you to make a decision, explain a mechanism, test an assumption, or try again without immediately showing the answer.
Vidbyte is still early, and I am actively trying to work out where this idea is most useful. How much intervention helps before it becomes irritating? Which kinds of questions produce genuine retrieval rather than another easy answer? What is the right balance between personalized guidance and the user's own effort? And how should an AI system measure progress when the real outcome is not a completed conversation, but something the person can do later on their own?
I would especially like feedback from Indie Hackers who are building products, learning difficult technical subjects, changing careers, or using AI heavily in their daily work. I am interested in the moments where AI has made you faster but not necessarily more capable, and in the workflows where you wish the tool would help you retain more of what you worked through.
Try Vidbyte here:
For developers interested in the agent infrastructure:
https://github.com/cerredz/Vidbyte-SDK
The learning skills are here:
Comment