Taimuraz Kaitmazov

Taimuraz Kaitmazov

I build deep, correct systems across unfamiliar stacks: AI platform, on-device NPU inference, backend, Linux internals.

I go deep in unfamiliar, hard systems most engineers won’t touch, and land results their own experts merge: AMD took 20+ of my fixes into their NPU toolchain, one into LLVM itself. AI does a lot of the typing. The discipline that keeps it correct (benchmarks, gold-sets, byte-exact diffs) is how I stay right in a stack I didn’t start in.

Lately

Claims with a public receipt open to it: click ▸ verify, or press v for all.

Built a from-scratch on-device NPU inference engine (open kernels, Apache-2.0). It runs model graphs on the laptop NPU: a 16-model parity loader, with ASR fully end-to-end and embeddings and vision also running on the device. Device trace showed it’s data-movement-bound (~93% of on-device time is overhead, not compute), so I cut it there. When AMD’s toolchain broke under the load, I fixed its compiler, operators, driver, and LLVM backend instead of stopping: 20+ fixes merged by AMD, one merged into upstream LLVM itself, 4 more in review.

The engine is public (Apache-2.0), and so are the PRs: public, third-party-accepted evidence of the work it required. Every fix exists because my own build hit it: the compiler (mlir-aie), the operators (IRON), the runtime (mlir-air), and the LLVM backend (llvm-aie).

xdna-engine (repo) · mlir-aie #3178 · amd/IRON #124 · llvm-aie #1059 · → full case

Fixes merged upstream in Zed and helm-diff; an open speculative-decoding PR to vLLM.

Zed & helm-diff read “Merged”; the vLLM PR is still open. Third parties accepted the code, or are reviewing it: not me.

zed #13840 · helm-diff #294 · vllm #44698

Production AI for a top-3 B2B accounting-software publisher: a payroll-compliance agent, a self-hosted OCR→LLM extraction pipeline (chosen on a labeled gold-set), vLLM-on-H100.

private

Open source

kde-mcp — computer-use MCP server for the Linux desktop (KDE/Wayland): accessibility-tree-first, not pixel-clicks. · read the case

tnl — self-hosted ngrok alternative: one-command reverse tunnels behind your own Caddy.

evbridge — evdev→Wayland input bridge for headless / nested compositors.

run-script-webpack-plugin — in NestJS’ official HMR docs; ~600K npm downloads/mo.

rethinkdb-ts — top contributor (~63% of commits, 9+ merged PRs incl. releases).

Contact

taimuraz@kaitmazov.com · github.com/atassis · linkedin.com/in/atassis

Remote-first · open to relocation · Moscow (UTC+3)