MANY FIELDS. ONE CURIOUS MIND.
An engineer.
More than one field.
Autonomous vehicles, learning systems, and the things that happen between disciplines.
Browse the indexPick a field.
See where it takes you.
7 fields and 59 project families. Spatial journeys are optional. Turn motion off above to navigate directly.
THE PROJECT INDEX
All projects, organized by field of study.
Within each field, projects are ordered by engineering depth, scope of Donald’s contribution, strength of the available evidence, and how clearly they show a distinct capability. This is an editorial judgment, not a performance score. Related repositories are grouped into project families.
Localization, navigation, vehicle control, and the physical consequences of a decision.
A ROS2/Gazebo course project connecting lane segmentation, curvature-aware speed control, and LiDAR/GPS localization.
START HEREThe broadest autonomy work: perception, control, localization, and explicit evaluation in one course sequence.
Real-time stabilization of an inherently unstable pendulum using a reaction wheel.
A potential-field controller evaluated over every goal assignment for five robots in a fixed obstacle-filled circular workspace.
A camera-to-control pipeline for a F1Tenth ROS2 platform, with confidence-aware path selection and explicit stop behavior when vision has no usable path.
Course exercises for a UR3 arm: recursive Tower of Hanoi planning, suction and motion coordination, plus an unfinished vision-based block-manipulation extension.
Driver monitoring, constrained-device inference, and reconstructing the world from images.
On-device driver-state monitoring developed during Donald’s research-engineering internship at Shanghai Ben’an Intelligent: eight detector classes, vehicle-aware alerts, and replay-driven validation.
START HEREAn end-to-end internship engineering story: on-device deployment, context-aware alerts, calibration traceability, and a 616-test validation suite.
A C inference pipeline for constrained hardware: sparse face detection, tracked regions, lightweight landmarks, and per-stage timing.
A local adaptation of VGGT that filters sky regions from reconstructed point clouds and adds reconstruction inspection tools.
Prompt-based image-segmentation experiments in Meta’s SAM 3 notebook, with local environment and prompt changes.
Forecasting, representation learning, reinforcement learning, and experiments inside the model.
Time-series foundation models meet nearest-neighbor retrieval, causal evaluation, and ablations that expose prediction-to-decision failures.
START HEREThe deepest experimental thread: model adaptation, retrieval, ablations, and useful negative results.
An extensive adaptation of the upstream Momentum Transformer research, including a PyTorch implementation and portfolio/execution reinforcement-learning layers.
A NumPy-only neural-network training library with dynamic computation graphs, reverse-mode differentiation and a small MNIST classifier.
Recurrent PPO, portfolio environments, and related regime and fundamental-overlay studies. Each experiment retains its own evaluation boundary.
Two iterations of a few-shot forecasting reproduction, including a tensor-level correction to what attention values remember.
A sequence of course studies in latent-variable generation, DDPM training and classifier-free guidance, then diffusion-based image deblurring.
Related DoubleAdapt/StockMixer, DeePM, and imitation-learning experiments in adapting models to A-share portfolio decisions.
An industry-aware equity research framework whose concrete implementation centers on ingestion, feature/label contracts, purged cross-validation and supervised model training.
An adapted AlphaPROBE research workspace with local point-in-time feature infrastructure and alpha-allocation additions.
Chart-image CNN/ViT, market-guided attention, and cross-sectional representation studies in one model-research family.
Graph, GNN, and HMM experiments investigate which stocks recover after a market leader’s limit-up streak ends.
A modular reproduction of the PyTorch CIFAR-10 tutorial that turns seed, augmentation, epoch and channel-width changes into selectable experiment configurations.
CAMEF and the Kalshi AI bot: upstream codebases retained as reference foundations, with original authorship clearly distinguished.
Running models locally, adapting them, and giving agent workflows an accountable structure.
Self-hosted Qwen workflows with controlled request budgets, reproducible decoding, structured decisions, and explicit inference-failure handling.
START HEREConcrete inference engineering across projects: budgets, structured outputs, runtime controls, and recovery.
An AI-work coordinator built around Git, markdown work nodes, a small state machine, and a CLI that owns mutations.
A local adaptation of an upstream financial time-series LLM pipeline to Qwen3.5, spanning inference compatibility and staged GRPO/LoRA/SFT training code.
A local-Qwen decision pipeline asks whether the real-world event behind a prediction market is likely to end within a chosen horizon.
A-share research workflows that turn textual evidence into structured long-only decisions, with accounting and governance checks.
A reusable agent workflow that separates planning, interface skeletons, implementation, and integration through typed patch artifacts and deterministic hooks.
An installation and activation layer combining upstream coding skills, specialized agents, and hooks into a reusable project toolkit.
A provider-neutral task-control architecture studied alongside the later ZhengMindYAGNI implementation.
An upstream reference for symbol-aware code retrieval and editing through language servers and MCP.
Industrial SOP verification, concurrency runtimes, and tools that turn a workflow into a system.
A Unix-like kernel in C and RISC-V assembly, from virtual memory to an interactive shell.
START HEREA complete operating-system stack: memory, processes, devices, storage, and a shell, implemented together.
An assembly-inspection prototype that tracks procedure steps, flags skips and ordering errors, and attaches video evidence for review.
Cooperative scheduling, epoll-backed I/O, timers, and synchronization in C++17. Waiting tasks yield instead of blocking a worker.
A local ApplyPilot adaptation adding a Claude CLI provider and crash-resilient parallel scoring.
Alternative native-authentication implementations inside a supplied React/GraphQL invoice application, with JWT access tokens, rotating refresh sessions and cookie-based browser integration.
Introductory CUDA course material for compiling a supplied device-query program and running it through Slurm.
A local collection of blockchain course slides and notes.
Execution, statistical arbitrage, market mechanisms, and research that makes its assumptions visible.
Order reconciliation, execution controls, queue-aware research, and an independently packaged refactor of a prediction-market system.
START HEREThe strongest execution-systems evidence: explicit order lifecycles, failure handling, and independent verification records.
Graph-conditioned residuals, pair selection, and a constrained search harness for testing alternative research modules.
Options research and execution tooling spanning implied-versus-realized volatility, short-strangle protocols, and an earnings-event variance decomposition study.
LPPLS bubble detection, an inspectable research cockpit, and related market-mechanism studies with frozen-history checks and rollback controls.
An integrated paper-research stack connecting a minute-data lab, portfolio optimizer, Nautilus parity backtest, observability and an agent control room.
A distinct event-market study combining broadcaster transcripts, settlement labels, crew effects and mention-time approximations.
Asset-pricing factor replication, index-enhancement experiments, and related upstream data infrastructure.
A short-horizon reversal system with clustering, market-specific tradeability, portfolio construction and futures hedging.
A multi-source signal-ingestion, replay and alerting layer for prediction-market research.
A repeatable multi-fold evaluation and dashboard system for technical signals on US minute bars.
Studies of market bottoms, public flow disclosures, IPOs, and observable policy-support signals, with rejected hypotheses retained.
Research notes that turn trading literature and informal claims into testable rules, experiment plans, and critical assessments.
A soccer pitch, a piano roll, a knowledge graph, and a study of technology beyond the code.
A real-time multiplayer math arena, built and launched in 72 hours.
START HEREA complete multiplayer product, from mathematical validation to live sessions, authentication, and rankings.
A Chinese soccer tactics teaching application with 52 situational challenges, 11 roles, three viewpoints and a 90-minute training plan containing 43 segments.
A Chinese-first reading and exploration application that turns video transcripts into staged summaries, source-linked ideas, and creator-level knowledge graphs.
A web pipeline converting YouTube/Bilibili piano audio into editable MIDI, with optional timing/dynamics cleanup and a saved library.
Two related Vue/CodeMirror implementations of a browser-only JavaScript/Python editor with execution, console output, saved files and export.
A dockerized live-information interface combining news signals, prediction-market context, public trade activity, and ADS-B flight overlays.
Local adaptations to an upstream subtitle-first transcription and summarization application, including a different model client, GPU transcription and platform-specific retrieval handling.
This site: a field-based project catalog and a local-Qwen guide that answers from reviewed claims and linked evidence.
A candidate-attributed course essay examines American Graffiti through histories of mobility, vehicle identity and the social consequences of transport technology.
BETWEEN THE FIELDS
ROBOTICS × EMBEDDED VISION
A lane mask becomes a vehicle-control input. A face landmark becomes part of a time-dependent alert. In both projects, the interesting engineering is how perception turns into a decision—and what happens when that input is unreliable.
EDITORIAL CONNECTION · GROUNDED IN THE PROJECTS