Amr Mareh · The Journal

Part II — Research log

The Research Log

This is where I’m tracking every piece of research behind Ulter: what it is, why I believe it matters, and how it moves the company’s goal forward. Each entry follows the same shape — what it is, status, why it matters — so the log stays easy to scan even as it grows.

The posts below are ordered to follow the dependency chain rather than the order I started them: Okran.ai the platform, then the Okran LLM we’re building toward, then Project Quantum Reduction that quantifies it, then ElaNa that runs the reduced model offline. Each step exists because of the one after it.

  • Entry 01 · Active development

    Okran.ai — The Multimodel Platform

    Ulter’s multimodel platform — all the frontier models in one place, with an autorouting layer that sends each request to the model that actually fits it.

    Why it matters: it answers the question that started Ulter — one place for all the frontier models — and it is the for-profit side of the company that funds the research below it.

    Read the entry →

  • Entry 02 · In development

    The Okran LLM

    Ulter’s own large language model — the model the platform is building toward, rather than one more model routed to from somewhere else.

    Why it matters: owning the model is what puts the decisions back in our hands — how it behaves, what it costs, and how small it can get.

    Read the entry →

  • Entry 03 · Research memo drafted

    Project Quantum Reduction

    Quantifying and fine-tuning the Okran LLM using quantum-inspired tensor network compression, in the style of CompactifAI / KARIPAP-type approaches — representing the weights as a network of smaller, interconnected tensors rather than one enormous matrix.

    Why it matters: it is the bridge between the Okran LLM and ElaNa. Without it, the offline model stays a wish rather than a product.

    Read the entry →

  • Entry 04 · In development

    Project ElaNa — The Offline Model

    Ulter’s first offline model: the Okran LLM shrunk by 95–99% so it can run without a live connection, and sync compressed weights down from the server whenever one is available.

    Why it matters: an AI that only works when you’re online isn’t actually accessible — and for some users, running fully offline is the requirement, not a nice-to-have.

    Read the entry →

  • Entry 05 · Coming

    Coming Research

    Reserved for what comes next — thesis chapters, papers I’m reading, and my own experiments, each written up in the same What it is / Status / Why it matters format. This log is a growing list, not a fixed set.