I’m building Lernex and training Metis-1.6 from scratch — AMA
Hi Product Hunt, I’m Giulianno, founder of Lernex.
We are building a learning system around a simple promise: make learning feel like it was made for you.
A learner can begin with almost anything: a PDF, notes, a recording, a screenshot, a question, or a topic, and turn it into the format they need.
Underneath that, the system is meant to learn from behavior: which explanations work, where difficulty should move, when a visual or interactive question would help, and what should be revisited next.
We are also training the Metis model family from scratch specifically for Lernex rather than fine-tuning a third-party checkpoint. The current work is Metis-1.6. The goal is product-level control over learning quality, latency, cost, structured generation, and persistent context.
Lernex has not had its Product Hunt launch yet. I’m here early because I want to build relationships, share the real work, and learn from the community before launch day rather than appearing only when I need upvotes.
Ask me anything about adaptive-learning product design, building persistent learner context, training a model from scratch, model/product tradeoffs, early user research, or what we still have not solved. I’ll answer candidly.
Replies
training a model family from scratch for persistent context specifically is a bold call this early, most teams default to fine-tuning until forced off it. what was the first tradeoff that made a third-party checkpoint feel like it was capping you, latency, cost, or something in how persistent context actually needs to work?
@sabber_ahamed
That was exactly the tradeoff: the limitation wasn’t only at the output layer.
We could fine-tune an existing model on Lernex’s schemas and probably improve its outputs, but it would inherit the same underlying latency, inference cost, context behavior, and capabilities of its base model. Proprietary APIs gave us strong intelligence but were often too slow and expensive for continuous personalization. Open-model providers improved the economics, but we still had to choose between speed, cost, and capability.
So we asked: why keep optimizing around compromises instead of designing for the actual problem?
Metis is trained from scratch and open-sourced on Hugging Face. That freedom lets us explore efficient architectures larger labs may consider too experimental. With Metis-1.6, we’re also preparing a paper on our new MoRE architecture, Mixture of Recursive Experts.
Without getting too technical, MoRE combines sparse expert selection with recursive computation. The goal is to give the model substantially more reasoning capability than its raw parameter count would suggest while keeping inference fast and inexpensive because only a fraction of the model is active at once. We’ll publish the architecture, evaluations, and limitations so those claims can be judged from evidence.
More importantly, Metis is trained around the work Lernex actually needs: long-lived learner context, large amounts of personal learning information, structured educational outputs, and decisions about what to explain, revisit, simplify, visualize, or test next.
Models from OpenAI, Anthropic, DeepSeek, and others are excellent general-purpose conversational models. They were built to be broadly useful, not specifically to teach one person continuously. We’re not trying to replace them at what they do best. We’re building for the problem they were never specifically competing to solve.