Aiden Kim

KAIST Aerospace Engineering · MLAI Lab · Agent Infrastructure

Website · Email · X

I work on ML systems: sparse Mixture-of-Experts inference and LLM serving on Intel Gaudi and Apple Silicon. I also build tools for AI agents and have a background in aerospace guidance and control. I am an Aerospace Engineering undergraduate at KAIST, expected to graduate in February 2028.

Now & recent

• KAIST MLAI Lab — Undergraduate researcher under Prof. Sung Ju Hwang, working on efficient MoE inference (arXiv 2610.37386) and heterogeneous LLM serving.
• DeepAuto.ai (Aug–Sep 2026) — Full-time research engineer on a project basis; independently brought up Kimi K3 serving on Intel Gaudi 2 across 16 HPUs.
• Pathtent — Co-founder & CTO, building a patent drafting and prior-art platform.
• Airbus Fly Your Ideas — Led In Vitro Firmware Red-Teaming. We advanced to the mentored Round 2 from 360+ teams and did not advance to the three-team final.

Selected work

• calibration-free-moe — Code for Channel Selection for Mixture-of-Experts Decoding Needs No Calibration Data (vLLM + Triton).
• recall — Semantic memory for AI agents without carrying full session history.
• transcribe-cli — Local document-to-Markdown extraction for PDFs, Office files, images, and more.
• code-intel — Lightweight code navigation and intelligence for agent workflows.
• opencode-context-compress — User-controlled context compression for long-running OpenCode sessions.
• GraduateKAIST — Private, client-side graduation planning for KAIST students.
• PINN_Guidance — Physics-informed neural networks for real-time aerospace guidance.

See the full project index →

Research

• Calibration-free MoE activation sparsity · first author · arXiv 2610.37386 · 2026

  Keeps 98.1% of full-model accuracy while reading 17.8% of routed-expert weights; MoE blocks run 1.60–1.90× faster than optimized vLLM at batch 1.

• Kimi K3 on Intel Gaudi 2 · DeepAuto.ai · Aug–Sep 2026

  Brought up a 2.8-trillion-parameter MoE model on Gaudi 2 from scratch across 16 accelerators: 8.7 tok/s single-user decode, 26.6× faster MoE prefill.

• Heterogeneous LLM serving · KAIST MLAI Lab · 2026

  DGX Spark prefill → Mac Studio MLX decode disaggregation in SGLang: +38.3% aggregate output throughput and −34.8% mean end-to-end latency versus one DGX Spark on Qwen3.6-35B-A3B.

• Long-context attention behavior · 2026

  Pilot experiments around RoPE degradation, attention sinks, and position addressability.

• Agentic completion reliability · 2026

  Studying whether intent-grounded, trajectory-aware stop gates can reduce false completion in long-horizon coding agents.

• Aerospace guidance · 2024—2026

  Missile range estimation, cooperative interception, and PINN-based trajectory generation; the PINN track is currently on hold.

_{Busan → Daejeon · KAIST Aerospace Engineering}