Senior backend engineer — 8+ years in production.

• Node.js / TypeScript (NestJS, Moleculer) — REST APIs from scratch, API gateways, message brokers, payment & telephony integrations, legacy refactors at ~20k LOC scale.
• Python (FastAPI, Django) — email pipelines on APScheduler + SMTP, document generation (PDF/DOCX) at template-set scale.
• Rust — current focus: on-device speech AI and agentic developer tools.

Local-first by default: zero cloud APIs, zero vendor lock-in, models that ship inside the binary.

Currently shipping: oh-my-kimi — wire-first Rust agent runtime · localmt — offline Android translation SDK targeting Redmi Note 14 airplane-mode proof.

Russian speech recognition on M1 in 700 ms. 11.4% WER. One binary. No cloud.

cargo install gigastt && gigastt serve
# WebSocket: ws://127.0.0.1:9876/v1/ws
# REST API:  http://127.0.0.1:9876/v1/transcribe

GigaAM v3 + ONNX Runtime · INT8 with 0% accuracy loss · WebSocket streaming + REST · Homebrew tap · CoreML / CUDA / CPU · v2.0.3

Speech AI — Rust + ONNX Runtime, fully offline

• polyvoice — Speaker diarization without Python · DER ~14% VoxConverse, ~23% AMI · 10× realtime on CPU, ~80% of pyannote accuracy
• phonex — Generic on-device STT engine · 10+ languages, ~70 ms / 5 s clip, single binary
• nihostt — Japanese STT (ReazonSpeech-k2-v2) · CER ~1.1% (clean) / 8% (full 309-clip bench) · ~200 ms latency, INT8 ~155 MB
• phostt — Vietnamese STT (Zipformer-vi RNN-T) · ~75 MB model · crates.io + pypi
• localmt — Offline Android translation SDK · GGUF + llama.cpp via JNI, arm64-v8a target

Agentic developer tools — Kimi ecosystem

• oh-my-kimi — Wire-first orchestration for Kimi CLI · scheduler-backed teams, ownership conflict detection, verification gates, proof/failure artifacts
• cargo-kimi — Cargo subcommand that scores Rust files 0–100 on contract quality (Hoare triples, panic safety, typestate, size, Result discipline) with LSP server · on crates.io
• kimi-guidelines — Composable configs, instructions, and skills for Kimi K2.6