Benchmark design:

• 56 execution-verified tasks spanning 14 scenarios and four quoting-hazard tiers
• Three transport contracts: raw shell, JSON tool calls, and wrapped bash -c
• Single-command oracles and state-based validators distinguish syntax failures from silent wrong results

Highlights:

• A catalog of 16 AI-coding tells across 5 tiers — duplication / reinvention, speculative abstraction, defensive slop, noise, and test slop — each with detection signals and before/after examples
• Three iron rules the catalog hangs off: behavior preservation is absolute (including error types), no tests → no edits, and one pattern-class per commit with the suite green after each
• A false-positive guard — severity 1 (“present but justified”: documented fallbacks, trust-boundary defenses, plugin registries) stays untouched; the goal is a healthier repo, not a body count
• First field test on a 13.4k-LOC agent-written ML repo: 24 findings, almost all Tier-1 duplication — pasted copies that had already silently drifted

Key Features:

• LLM front-end + deterministic engine: free-form input → LLM intent parser → canonical action → original Adventure engine → state transition → LLM narrator
• The game kernel owns all world state (rooms, items, puzzles, death, score) — the LLM never invents state, so there is no hallucination drift
• Natural-language play in Chinese or English — no more memorizing the exact verb (look, not see)
• LLM narrator turns terse engine output into modern, readable prose and explains failures instead of barking “I don't know that word”
• Brings the 1977 Colossal Cave Adventure to modern players while fully preserving its rigorous puzzle world model

Setup:

• Compares three system-prompt modes: base (English), zh_compact (Simplified Chinese), wy (Classical Chinese)
• Benchmarks: IFEval, MATH-500, MMLU-Pro; fixed 2048-token output budget to rule out truncation artifacts; per-run token accounting
• Manifest-driven pipeline over both API models and local Qwen models

Preliminary findings:

• zh_compact is the most robust default across model families
• Classical Chinese shows only conditional gains — strongest on gpt-5.4, but no stable advantage on qwen3-4b

Key Features:

• End-to-end pipeline: audio capture, VAD, ASR model inference, and system integration
• Lightweight on-device post-processing models for punctuation restoration and text formatting
• Speech editing mode: real-time correction and insertion with voice commands
• Native IBus and Fcitx 5 front ends for system-wide input across GNOME, KDE, and other Linux desktops
• Optimized for low-latency offline operation without cloud dependencies

My Contributions:

• Co-designed the DFSPH coupling that enforces incompressibility (constant-density & divergence-free constraints) on Gaussian particles each step
• Implemented uniform internal filling — converts surface-biased 3DGS into SPH-ready volumes via a smoothed opacity field over a uniform 3D grid
• Implemented an SPH ∇v–based implicit covariance update so anisotropic Gaussian shapes align with local flow during simulation
• Compared against PhysGaussian (MPM) on Synthetic-NeRF scenes (Materials, Hotdog, Ficus) — preserves liquid-like coherence and density with a single point-based representation

My Contributions:

• Two-stage training: linear probing on frozen MAE features (8 epochs) followed by end-to-end fine-tuning at reduced lr (25 epochs)
• Class-imbalance handling via weighted random sampling (inverse class frequency) and AdamW with weight decay; automatic mixed precision for stability
• Post-hoc decision calibration: prior adjustment (logits − τ·log prior) plus class-specific bias to suppress systematic nevus over-prediction
• Standard dermatology augmentations (random crops, ±30° rotation, color jitter) on 224×224 normalized inputs

My Contributions:

• Implemented multi-method point cloud denoising pipeline (statistical outlier removal, voxel downsampling)
• Integrated ICP-based pose refinement for robust localization under sensor noise
• Improved robustness for mobile manipulation in dynamic environments

My Contributions:

• Reimplemented PyTorch-style autograd with computational graph construction and reverse-mode differentiation
• Built fully connected neural network layers, activation functions, and optimizers (SGD, Adam)
• Validated against standard benchmarks to ensure numerical correctness

My Contributions:

• NJUOS lab series (Spring 2023): process scheduling (round-robin, priority-based), virtual memory management (paging, demand paging)
• MIT xv6 lab series (Summer 2024): inode-based file system, basic POSIX system calls, additional kernel subsystems
• Hands-on practice across boot, scheduler, memory, and storage layers of a teaching kernel

My Contributions:

• Fine-tuned LLaMA model for domain-specific Q&A tasks
• Implemented vector database indexing pipeline for efficient document retrieval
• Built Vue.js frontend and Spring Boot backend with RESTful APIs