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

email · linkedin · kaelvalen.vercel.app

My main project right now is trainscope, a post-mortem debugger for LLM training loss spikes. It started as a tool I needed during ENGRAM runs and has since become the focus: a CUSUM-based change-point detector catches gradual loss drift 5-20 steps before it turns into a full spike, validated with false-positive rate testing across 280+ held-out noise scenarios (0.0%) and sensitivity testing across drift magnitudes from 0.10σ to 0.50σ (100% detection). When a spike hits, the UI reconstructs the chronological failure cascade: which layer moved first, whether it was a distributional shift or a gradient explosion, and how many steps of warning were available before collapse.

Before this, I closed out NOESIS, a continual-learning system that tried to inject retrieved memory (via a Titans-style associative matrix + test-time training) directly into a frozen LLM's hidden states. The result was negative and worth stating plainly: a frozen backbone's hidden manifold cannot be given new semantic content by an untrained vector injection, no matter how carefully the geometry is aligned (verified across KV-cache-purge controls, negative controls, and a 50-fact automated benchmark with 0% recall). Full writeup in FINDINGS.md.

• trainscope: see above. Post-mortem debugger for LLM training loss spikes. FastAPI + React, CUSUM change-point detection, WandB auto-attach, on PyPI.
• noesis: continual-learning experiment with frozen backbone + vector memory injection. Concluded negative result, documented in FINDINGS.md.
• engram: hybrid linear-recurrent backbone with Mamba-2 SSD blocks interleaved 3:1 with Gated Delta Rule blocks. Applied to PTB-XL, Speech Commands, and sCIFAR-10. Pipeline validated end-to-end.
• connor: terminal-native CI/CD runner in Rust. TOML config, DAG-parallel execution, no server.
• latch-lang: programming language written in Rust. On crates.io.
• weave: local-first, plugin-based productivity app. Tauri + React + Rust.
• nanonet: distributed service monitoring platform. Go backend, Rust agents, React frontend (~70k lines). Archived.
• beyond_transformer: PULSE, predecessor to ENGRAM. Kept public as design record.

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.