I'm Minjoon Choi, a second-year Electrical and Electronic Engineering student at Hongik University, interested in on-device AI — running models fast and reliably on the user's own device, without a GPU server or the cloud. Research interest: low-bit quantization and model compression for on-device inference.

Two quantization experiments (int8-quant-lab, lowbit-qer-lab): the plan and the honesty rules are mine; implementation, measurement and documentation were done by Claude Code.

Besides those, I've built three offline tools that run on a single laptop, all made so that an AI assistant could read my own textbooks, lectures and exam papers. Design decisions in them were settled by measured A/B comparisons. The code was written with an AI coding assistant (Claude Code); problem definition, measurement, and the decision to adopt each design are mine.

• int8-quant-lab — TFLite/LiteRT INT8 quantization re-implemented in NumPy and matched bit-for-bit: every scale, zero-point and int8 value the TF 2.21 converter produced for the MLPerf Tiny KWS model, and the integer outputs of three execution paths (reference, optimized, XNNPACK) on the full test set, which round in six different ways across paths and platforms. FP32 92.17% → INT8 92.33% (statistically indistinguishable); 4-bit per-tensor symmetric weights collapse to 57.4% (simulated) — and the analysis shows the weight error is exactly what the Δ²/12 model predicts; the damage comes from large weights on a coarse grid and from one layer's error being amplified on the way to the logits. Tools: NumPy, LiteRT, TensorFlow 2.21 (conversion), flatbuffers, GitHub Actions (Linux/Windows).
• lowbit-qer-lab — low-bit weight quantization with low-rank error reconstruction (W ≈ Q + LR) in NumPy: group-128 RTN, QER, a preserve-then-quantize rank split and a surrogate for choosing the split, compared on 49 real linear layers of SmolLM2-135M and TinyLlama-1.1B. The split's gain over QER in weight error is set by a single number, ρ_r(W) (correlation −0.97/−0.98; first-layer q/k projections gain 41–61%, TinyLlama layer-0 q_proj 0.247 → 0.101), the optimum is almost always k = r, and surrogate regret is ≤ 3.9%. Perplexity on WikiText-2 moves the other way: the split has the lower weight error yet QER is far better (3-bit 80.9 vs 121.1), and per-layer output error with real activations does not explain it — an open question; at equal bits per weight, one more bit of RTN beats both (4-bit RTN 22.7). Not a paper reproduction. 59 tests. Tools: NumPy, SVD, a torch-free safetensors reader, PyTorch/transformers.
• pdf-ocr-korean-textbook — Korean textbook PDFs to Markdown with LaTeX math. Rebuilt the formula recognizer with a KV cache and int8 quantization by transplanting its ONNX weights into a transformers model (4.16× faster recognition); encoder on the iGPU via DirectML (1.37× end-to-end). 7,530 pages from 9 books, 93.8% character accuracy / 96.5% word recall against hand-transcribed pages. Tools: ONNX Runtime (DirectML), PyTorch, transformers, optimum, Tesseract, pdfium.
• lecture-transcriber — recorded lectures to timestamped Markdown that includes what was on screen. CPU-only int8 Whisper; slide changes caught 48/48 within ±1 s on a synthetic lecture; silent omissions turned into marked ones. Tools: faster-whisper (CTranslate2 int8), ffmpeg, Tesseract.
• hwp-to-pdf — Hangul word-processor files (.hwp/.hwpx) to PDF without Hancom Office. Repaired the open-source renderer rhwp's broken text layer (100% Hangul across four extractors) and, by comparing against Hancom's own renderings, fixed eight layout bugs in rhwp (page-1 silhouette agreement on real exam papers 79.9% → 98.6%; all reported upstream). Each conversion is self-checked against the preview image Hancom stores in the file. Tools: rhwp (Rust, patched and built from source), pikepdf, pypdfium2.

Contact: mijnch@gmail.com