Projects
Research Prototypes & Systems
Projects we have implemented, reproduced, and measured—documented with demos, evidence, negative results, engineering boundaries, and next steps.
2026Reproducible experimental system · pipeline audited
Offline recommendation gains may create concentrated exposure, overloaded merchants, courier imbalance, and delivery delays downstream.
We connect Top-K recommendation to merchant exposure, courier matching, and peak-hour simulation, measuring accuracy, fairness, ETA, timeout rate, and platform utility together.
- 0.4509
- Seq-Tuned Recall@20
- 24.98 min
- KG + Batch mean ETA
- 0.5697
- best mean platform utility
- Python
- Pandas
- scikit-learn
- Streamlit
- Plotly
- NetworkX
- pytest
View approach and outcome ↗2026Active research · core reproduction complete
Multimodal models may miss diagram relations or invent invalid constructions, while apparent gains can be confounded by extra reasoning tokens and output formatting.
We turn auxiliary geometry constructions into executable, verifiable, and replayable intermediate states, then use paired experiments to separate visual gains from reasoning-protocol gains.
- 149 passed
- core and web tests
- 85 / 92
- Flash confirm correct
- +2.000
- two-stage utility gain
- Python
- FastAPI
- Next.js
- React
- SVG
- SQLite
- pytest
View approach and outcome ↗2026Manuscript in preparation · core modules evaluated
Ideas describe goals and hypotheses while datasets describe fields, modalities, and empirical boundaries; keyword search cannot reliably align the two.
We build a shared Idea–Dataset research memory for both idea-to-dataset retrieval and dataset-grounded research ideation.
- 217,942
- aligned Idea–Dataset pairs
- 64,252
- source papers
- 0.3334
- best relaxed MRR
- Python
- Sentence-BERT
- Faiss
- RAG
- LLM
- LaTeX
- pytest
View approach and outcome ↗2025–2026Usable prototype · engineering boundaries audited
Code, logs, and diffs already live in the shell; repeatedly copying them into a web chat breaks the natural command-line workflow.
I combine one-shot prompts, local conversation history, system prompts, R1 reasoning streams, and Unix pipes in a Rust CLI, with an explicit audit of its current boundaries.
- 5.7 MiB
- arm64 release binary
- 3 modes
- stateless/new/resume
- 0 tests
- current repository tests
- Rust
- Tokio
- Reqwest
- clap
- Serde
- TOML
- rust-i18n
View approach and outcome ↗2025Engineering prototype · core flow demonstrable
Computing labs need reproducible per-student environments and centralized management of quotas, assignments, spaces, and cluster resources.
We combine browser IDEs, student isolation, time quotas, and cluster scheduling into a multi-user cloud development platform prototype, while documenting unfinished coursework, storage-quota, and production work.
- 7/7
- focused student API tests
- 1486
- student build modules
- 1526
- admin build modules
- Vue 3
- TypeScript
- Python
- Flask-OpenAPI3
- Redis
- asyncio
- Kubernetes
View approach and outcome ↗
