# Cerin Amroth — ML, JWST & Security Research

Source: https://cerinamroth.com/

[Home](https://cerinamroth.com/)[ML](https://cerinamroth.com/ml/)[JWST](https://cerinamroth.com/jwst/)[Security](https://cerinamroth.com/policy/)[Contact](https://cerinamroth.com/#contact) 

# Cerin Amroth — independent research: ML-systems efficiency, JWST observational work, and software security

      

Large models on smaller machines. Public telescope images, examined closely. Software claims, put to the test.

  

 

## ML systems — large models, smaller machines

 

 [ Loggetta Check your hardware, plan a training run, and keep the results. One install includes the runtime and kernels. v0.7.0 · Plan and run training · `pip install loggetta` → ](https://cerinamroth.com/ml/) [ experts4bit-qlora Fine-tune and serve large MoE models on one GPU. Move weights to RAM or SSD when GPU memory runs out. v0.58.1 · Get started → ](https://cerinamroth.com/ml/experts4bit-qlora/) [ grouped-nf4-gemm Compute directly on packed 4-bit weights. GPU kernels, memory tools, and tests you can reproduce. v0.51.0 · Get started → ](https://cerinamroth.com/ml/grouped-nf4-gemm/) [ OpenVINO Upstream fixes keeping Intel-GPU LLM inference off the CPU-fallback path — three merged. ](https://cerinamroth.com/ml/openvino/) 

 

[Latest ML research · 8 October 2026](https://cerinamroth.com/research/#update-2026-10-08) · [GitHub: Loggetta](https://github.com/pjordanandrsn/loggetta) · [PyPI](https://pypi.org/project/loggetta/) · [models](https://pypi.org/project/loggetta/#models) · [measured speed and architecture](https://cerinamroth.com/ml/) · [lower-layer solution guides](https://cerinamroth.com/ml/solutions/) 

 

 

 

## Start with the problem

 

 | Problem | Route 
 | I want to plan and run model training | [Loggetta](https://cerinamroth.com/ml/) 
 | My MoE still OOMs after `load_in_4bit=True` | [experts4bit-qlora internals](https://cerinamroth.com/ml/solutions/bitsandbytes-moe-load-in-4bit-still-ooms/) 
 | I need to QLoRA-train fused MoE experts | [experts4bit-qlora](https://cerinamroth.com/ml/solutions/qlora-fused-moe-experts/) 
 | My model is larger than GPU or host memory | [e4b residency paths](https://cerinamroth.com/ml/solutions/run-moe-larger-than-vram/) 
 | Dequantizing every expert before GEMM is too slow | [grouped-nf4-gemm](https://cerinamroth.com/ml/solutions/nf4-grouped-gemm-without-bf16-materialization/) 
 | I need native MXFP4 or INT4 expert compute | [grouped-nf4-gemm](https://cerinamroth.com/ml/solutions/native-mxfp4-moe-inference/) 
 | I want to inspect the evidence methodology | [Receipts-driven engineering](https://cerinamroth.com/research/receipts-driven-engineering/) 

 

 

 

## Astronomy — explore real JWST observations

 

 [ Little red dots — an ongoing search Map published little red dots, explore an ongoing search of public JWST images, and follow individual sources through real telescope observations. ](https://cerinamroth.com/jwst/lrds/) 

 

 

 

## Security — find the flaw, help get it fixed

 

 [ Vulnerability research Good-faith research & coordinated disclosure. ](https://cerinamroth.com/policy/) [ Published disclosures ZisK zkVM: a soundness bug in signed remainder, fixed upstream. Vendor advisory GHSA-qjgg-fcj3-p9x3, rated Critical. ](https://cerinamroth.com/disclosures/) 

 

  

Contact: [jordan@cerinamroth.com](mailto:jordan@cerinamroth.com) · security reports: [security@cerinamroth.com](mailto:security@cerinamroth.com) · [security.txt](https://cerinamroth.com/.well-known/security.txt) · consulting: [jordananderson.work](https://jordananderson.work) · [Site data](https://cerinamroth.com/site.json)

 

Jordan Anderson’s [public GPG key](https://cerinamroth.com/.well-known/jordananderson-pubkey.asc) · Ed25519, expires 11 June 2028
Fingerprint: `DAEF C03B 9188 793C 80E5 20C8 68F0 6663 C7FB 250F`

  © 2026 Cerin Amroth Security Research, LLC · CERIN AMROTH SECURITY RESEARCH™