Skip to content
View kaelvalen's full-sized avatar
😺
😺

Highlights

  • Pro

Block or report kaelvalen

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
kaelvalen/README.md

Mehmet Arda Hakbilen (kael valen)

I work on efficient sequence architectures and the infrastructure they run on. Ankara, Türkiye.

email · linkedin · kaelvalen.vercel.app

Now

My main project is PRISM, a hybrid linear-recurrent backbone: Mamba-2-style SSD blocks interleaved with Gated Delta Rule blocks at 3:1, applied with identical hyperparameters to 12-lead ECG (PTB-XL), Speech Commands, and sequential CIFAR-10. The bet is that a single backbone, without language-specific design choices, can get within bootstrap CI of xresnet1d101 (~0.928 macro AUROC) on clinical time series.

Both mixers are written from scratch, with numerical-equivalence tests against torch.associative_scan and the FLA Triton kernels. 270+ tests in total (equivalence, fp64 gradcheck, streaming state-passing, property-based). The pipeline is validated end-to-end on an RTX 5060 laptop; the full paper matrix hasn't been run yet. The draft is aimed at the ICML 2026 ES-FoMo IV or NeurIPS 2026 ENLSP-VI workshops.

The current sub-project is MoM, in the same repo: rather than fixing the SSD:GDR ratio by hand, a per-token router chooses which memory primitive handles each token. The first spike gate failed. Routing doesn't collapse, but recall lags the GDR-only baseline, so I'm working through the mitigations. Everything has to fit in 8 GB of VRAM, which is a constraint I've come to appreciate.

Projects

  • prism — see above. PyTorch, active.
  • trainscope — flight recorder for LLM training loss spikes. FastAPI + React, Welford z-score detection, on PyPI. Wrote it because I needed it during PRISM runs.
  • connor — terminal-native CI/CD runner in Rust. TOML config, DAG-parallel execution, no server.
  • latch-lang — my own programming language. Rust, on crates.io.
  • weave — local-first, plugin-based productivity app. Tauri + React + Rust.
  • open-glyph — pixel-art editor for the Nothing Phone glyph matrix. Kotlin.
  • nanonet — monitoring platform for distributed services: Go backend, Rust agents, React frontend, ~70k lines, built solo. Mostly an exercise in shipping the boring parts too. Archived.
  • beyond_transformer — PULSE, the predecessor to PRISM. Kept public as a record of the design choices that led to the current one.

Stack

PyTorch for the ML work, with Triton when there's no way around it. Rust and Go for systems, TypeScript/React for frontends. My laptop runs NixOS, so a couple of repos here are just dotfiles.

GitHub metrics

Pinned Loading

  1. prism prism Public

    Modality-portable hybrid sequence backbone (SSD + Gated Delta Rule) — one architecture for ECG, audio, and images

    Python 1