I build research-shaped systems: online judges that understand teaching data, AI-agent infrastructure that can inspect and improve code, and memory-centric modeling experiments that try to make long-context reasoning cheaper, sharper, and more auditable.
My favorite work lives at the boundary where a mathematical claim has to become a running system: proofs, tests, architecture diagrams, telemetry, and a repo that another engineer can actually reproduce.
| Lane | What I am building | Public signal |
|---|---|---|
| Long-context AI research | RetNet-style retention, Engram lookup, Block Attention Residuals, milestone snapshots | engram-retention |
| AI-native education systems | Online judge platform for school teaching, evaluation, and AI-assisted governance | CodeNexus |
| Agent memory and evolution | Systems where agents learn, forget, route, and self-improve over time | dna-memory, evolver, mindx |
| Multi-agent learning | Interactive classrooms and coordination surfaces for AI-assisted learning | OpenMAIC |
| Edge and signal systems | WiFi sensing, inference pipelines, and non-visual perception systems | RuView |
|
PyTorch research scaffold for budgeted long-context memory: RetNet recurrence, hashed Engram lookup, Block Attention Residuals, and milestone snapshots.
|
An AI-native online judge platform designed for school teaching, judging, integrity workflows, and educational data intelligence.
|
flowchart LR
A["Research questions"] --> B["Formal assumptions"]
B --> C["Executable prototypes"]
C --> D["Tests and ablations"]
D --> E["Docs and proof trail"]
E --> B
C --> F["Education systems"]
C --> G["Agent infrastructure"]
C --> H["Memory architectures"]
C --> I["Signal and edge systems"]
- Start with the actual system, not a slogan.
- Keep claims narrow until tests make them stronger.
- Prefer architectures that can be inspected, reproduced, and falsified.
- Build AI features around evidence, governance, and workflows, not just chat.
- Treat documentation as part of the system, not packaging after the fact.
Public repos: 19
Recent center of gravity: Engram Retention, CodeNexus, OpenMAIC, RuView
Main languages by public repo count:
JavaScript #### 4
C++ ### 3
Python ### 3
Rust ### 3
TypeScript ### 3
Go # 1| Signal | Why it matters |
|---|---|
| Research code | engram-retention keeps proofs, configs, tests, and citation metadata together. |
| Systems code | CodeNexus pushes AI-native online judging toward real teaching workflows. |
| Agent code | dna-memory, evolver, and mindx explore memory, self-evolution, and digital identity. |
| Interface code | OpenMAIC experiments with multi-agent learning surfaces. |
This section intentionally avoids third-party dynamic stat renderers. Some
popular README stat services currently return 503 errors, which GitHub surfaces
as Error Fetching Resource.
Email: cachoidxx@gmail.com
GitHub: https://github.com/XXY-CH