I build local inference runtimes, study how language models behave, and test where evaluation pipelines lie.
Most of my work is Rust, quantized CPU inference, experiment plumbing, and Arabic morphology. I benchmark assumptions, chase failure modes through model files and kernels, and write down what broke.
01 / latest
Four useful signals from the lab, not the whole archive.
release candidate
Strict versioned experiment bundles, six semantic hook sites with exact restoration, a GPUI Kit console, and a reproducible golden ladder against the pinned llama.cpp reference.
research
Probe recovery is not native transfer. A continuation improved training fit while losing the only identity that actually transferred.
research
The gain was concentrated in number, and it moved when the representation interface or lexical split moved. Less headline; more useful.
failure mode
Fuzzing GGUF loaders made the boundary obvious: metadata controls execution, so parsing it is a security decision.
02 / systems
Mostly instruments for asking models uncomfortable, reproducible questions.
CPU-first Rust inference over GGUF: packed Q4/Q6 execution, hidden-state interventions, reproducible bundles, multimodal paths, and auditable tool traces.
Maps limitations, missing evaluations, and future-work claims from papers. Deterministic, evidence-first, and deliberately not a research-gap generator.
A small x86_64 kernel with paging, a heap, cooperative async execution, ramdisk, and a shell. Fewer layers; fewer places for the explanation to hide.
A local document-retrieval system for PDF collections with incremental indexing, context-aware chunks, embeddings, and local inference.
03 / research
Three questions currently shaping the work.
11-model audit
How much comes from morphology, and how much comes from lexical overlap, prompt leakage, token position, or a convenient pooling choice?
current answer →causal localization
The broad “Arabic degrades under Q4” story did not survive the tested matrix. Single-layer patches did: rare boundary failures localized to L7 in Qwen and L1 in Llama.
experiments →runtime forensics
A Q4 file can still become an f32 runtime. I care about resident representation, execution paths, and provenance, not just the filename.
autopsy →04 / field notes
Technical notes after the benchmark, bug, or methodological caveat became too annoying to leave undocumented.