SULTAANresearch notebook · .io

Vol. 1Phase 3: pretraining infrastructureRepository private

Research notebook · updated 5 October 2026

Notes from building a sovereign AI kernel

Sultaan is a research project to build a complete AI stack that runs on hardware you control: local inference, an agent loop, persistent memory, sandboxed tools and, eventually, a language model trained from scratch. The code lives in a private repository, so the engineering notes are published here.

Repository private · updates here

§ 0.1as of 4 Oct 2026

Status at a glance

Stage
Research prototype. Not a product, and not a dependency of any launch.
Phase
Phase 3, pretraining infrastructure, entered on 9 June 2026. The trainer is built and proven on CPU; a real run at tier T1 or above has not happened yet.
Model
None released. The largest recorded run is a 33.9M-parameter first-contact model trained on a 6 GB laptop GPU on 16 July 2026.
Code
About 18,000 lines (Python about 16,300, Rust about 1,400, shell about 360), 341 test functions and 26 architecture decision records on the main branch.
AI calls
No hosted model providers. Inference and training run on hardware you control; base models and datasets are downloaded from the Hugging Face Hub.
Licence
An Apache-2.0 placeholder, with a custom licence planned. No code or weights are publicly released.

noteFigures on this site come from the repository, its decision records and the portfolio review of 1 October 2026. Where a figure has no run record behind it, the page says so.

§ 0.2contents

What is in this notebook

  1. 01ArchitectureA Python and Rust monorepo with one seam per concern, plus the 26-entry decision log.
  2. 02Training approachA from-scratch decoder, a tier ladder from about 20M to 1.5B parameters, streamed data with provenance, and the runs recorded so far.
  3. 03Evaluation planPerplexity, benchmark subsets, hallucination probes, a grounding verifier and a promotion gate that is built to fail.
  4. 04RoadmapFive phases, the ten engineering sessions planned next, and what is deliberately out of scope.
  5. 05Follow progressA dated research log, an Atom feed and a contact route.
§ 0.3principles

Four working rules

  1. Own the stack. The toolkit runs on your own machine. It makes no calls to hosted model providers, and the memory layer carries a guard against sending data to cloud hosts.
  2. Small and owned beats big and borrowed. The agent loop, the evaluators and the training loop are written in-house. Any new dependency needs a written decision record first.
  3. Honest ceilings. The largest tier planned for a single GPU is about 1.5B parameters. Pretraining a 30B model from scratch is a multi-node cluster job, and the project never promises it on one machine.
  4. A gate that can fail. A checkpoint is promoted because it passes measurable checks, not because it is the newest. Skipped checks are always shown, never quietly counted as passes.

note“Sovereign” here means the model, the memory and the tools run where you decide, with no vendor in the loop.

§ 0.4position

The honest position

The portfolio plan treats Sultaan as research, not a launch item. A from-scratch model at the largest single-GPU tier would not match the hosted models that the portfolio’s products already use. If a “sovereign” model is wanted sooner, the practical route is LoRA fine-tuning of an open-weight model with the QLoRA scaffold that already exists in the repository.

Long term, a model tuned or trained here could become a local tier behind the shared AI gateway, which is part of the portfolio’s Shared AI Runtime & Data Plane. Its current status: Status unknown. That step is gated: it needs a run record showing the model beats the hosted low-cost tier on a named product task.

Prefer the plain-language overview? It is at sultaan.ai.