Why One Small Fixed Tree Suffices: The Cost Geometry of Draft Trees for a Multi-Token-Prediction Drafter
Gyeongmin Kim, Ayoung Moon, Seung Jin Lee. Under review, 2026.
Experience
NC AI — Data Engineering TeamMay 2026 – present · Seongnam, KR
LLM inference optimization & serving — MoE quantization (W8A16), speculative decoding for a data-generation pipeline, and a distributed vLLM serving gateway in Go.
Yonsei University Health System — Research EngineerMar 2025 – Oct 2025 · Seoul, KR
Led dev for NGS clinical report pipeline & SICU false-alarm monitoring desktop app.
NCSOFT — Audio Data TeamJul 2024 – Jan 2025 · Seongnam, KR
Built end-to-end audio post-processing automation for TTS data pipelines.
Taiyaki Studios — AI TeamJan 2023 – Jul 2023 · USA, Remote
Built a complete TTS training toolkit and production inference pipeline.
Personal Projects
Model Optimization (Nov 2024 – present)
Speech models re-implemented in dependency-free pure C, running in real time on-device.
nemotron-asr-streaming.c — 0.6B streaming ASR (FastConformer + RNN-T). Cache-aware streaming, W8A8 quantization with custom "Q8P" 4-row tile packing, NEON/AVX2 SIMD, up to 16 threads.
[ docs ][ quantized model ]
LILAC — Zero-shot real-time voice conversion from a 3s reference (OpenVoice v2 port). Streaming HiFi-GAN decoder, 2-thread SPSC audio pipeline, RNNoise SIMD. RTF 0.7–0.8 on CPU, 270–330 ms end-to-end.
[ optimizations ]
DeepFilterNet3.c.wasm — Noise-reduction model in pure C/WASM, real-time in mobile browsers. ~1 ms/frame on MacBook M2, ~4 ms on Galaxy S23.
[ optimizations ]
Diagnosed chunk-boundary artifacts as receptive-field truncation; removed them with overlap-save chunk inference (neighbor-chunk context, keep center) at no added latency.
Multilingual TTS (Oct 2022 – Feb 2025)
Data — BGM removal (Kim Vocal) → speaker diarization (pyannote) → VAD segmentation → Whisper transcription: ~30k clips (3–10 s).
Models — VITS derivatives: JA2ML-VITS (19-language speech from Japanese-only data), JK-VITS (KO/JA bilingual), RVC-VITS (RVC-converted training data).
novel-reader — Android app translating novels from 7 Japanese sites with a proper-noun dictionary system.
EhndWebTranslate — Async Japanese web page translator with real-time/document/novel modes.
UserDict4Papago — Proper-noun dictionary overlay for Papago KR-JP translation.
El Nino (Apr 2024 – Apr 2026)
Founded El Nino (university startup club → sole proprietorship) and built Knoc, a real-time translation subtitle service for classes and conferences — launched Mar 2026, covered by Multilingual.com.
Rewrote the Python prototype in Go for large-scale WebSocket handling; Redis Pub/Sub subtitle rooms, distributed architecture, user/admin dashboards with PAYG & subscription billing.