Selected projects
From-scratch engine (open kernels included), running whole model graphs natively on the laptop NPU via the open MLIR-AIE / IRON stack; a general open engine of this kind (vs isolated example kernels and closed-binary turnkey runtimes) didn’t exist. A 16-model / 11-architecture parity loader (BERT · Whisper/Parakeet/GigaAM · ViT/DINOv2/ResNet-18/CLIP · ESM-2), rel-err ~4e-3. ASR is fully end-to-end on the NPU; embeddings (BGE, 2.5–4× vs host) and vision also run there; the rest, including ESM-2, are in progress. Core finding: proved by device trace that it’s data-movement-bound. ~93% of on-device time is overhead, not compute. So I fused a 12-layer decode into one NPU dispatch (~33 host dispatches → 1 per token), weights + KV resident. Lower package power at equal accuracy (WER 0.117): an energy/offload win (RAPL package-level), not a latency claim. When the toolchain broke under the load I fixed AMD’s compiler, operators, driver, and LLVM backend rather than downscope. 20+ merged, one merged into upstream LLVM itself, 4 more in review. Public (Apache-2.0).
MLIR-AIE · IRON · Peano · XRT · Rust · sole architect · mlir-aie #3178↗ · amd/IRON #124↗ · llvm-aie #1059↗
AI payroll-compliance agent over 1С:ZUP
For a top-3 Russian B2B accounting-software publisher. Daily-reconciliation AI agent on the dominant Russian payroll/ERP platform + a Go/Rust MCP server; 30+ validation-rule registry, tax/labor-regulation cross-references, on-premise connector. Closed-beta with real client accountants surfaced specific compliance discrepancies in 10+-year ERP legacy.
Go · Rust · Anthropic SDK · sole architect · closed beta
Document-intelligence pipeline for accounting primaries
Self-hosted OCR → LLM extraction of structured fields from Russian accounting primaries, PII kept on-premise (dots.ocr on H100 + an LLM extraction stage). Built a small-scale eval harness: a 6–9-doc hand-labeled gold-set plus ~30 model-labeled “silver” docs, a human-in-the-loop labeler and a scorer. Then benchmarked OCR→LLM architectures on it, choosing the production design on ground truth rather than impression (privacy-vs-accuracy tradeoff quantified).
vLLM · H100 · FastAPI · sole engineer · production
The kind of tool Anthropic’s computer-use covers for macOS/Windows, still rare on Linux/Wayland. Built from scratch, solo. Drives apps through the accessibility tree (like a screen reader) rather than pixel-clicks: more reliable and safer. Safety policy gate, ADR-documented.
Rust · KDE Plasma 6 / Wayland · open-source · github.com/atassis/kde-mcp↗
Multi-tenant medical SaaS — OCR + LLM PII depersonalization
Local PaddleOCR-VL + Presidio + LLM-based PII detection for Russian PII (SNILS / INN / passport) with bounding boxes; per-tenant isolation (nginx routing) + a ClickHouse dedup buffer in front of Postgres.
PaddleOCR-VL · Presidio · Werf/Helm/K8s · designer & lead
Three-language MCP server stack for 1С
Coordinated Go + Rust + TypeScript/Bun stack exposing 1С enterprise data to agents over MCP, behind a unified backend-neutral query abstraction with pluggable backends. 15+ MCP tools.
Go · Rust · TS/Bun · sole architect · private
Headless-GPU containerization for remote graphics
A container that runs a full GPU graphics stack headless on modern Linux (Wayland, DRM/GBM, CUDA/DMA-BUF, seccomp isolation). Deliberately general: remote desktops, thin clients, and several isolated clients sharing one GPU, not just game streaming (its first use case). Built on current software (native Wayland) rather than the usual X11-era images.
Wayland · DRM/GBM · CUDA · seccomp · private
Drug-interaction API audit (medical)
Audited a production drug-interaction service against multiple authoritative sources and found it flagged only ~52.7% of the interactions it should have. A fast, decisive evaluation that resolved a task another team had left open for 2–3 months.
evaluation / audit · sole · unblocked a cross-team stall