Systems research · Resource optimization
The systems problems I enjoy most begin after something works: can it use fewer resources, stay responsive under pressure, and degrade gracefully when resources run out?
Currently, I study Linux memory pressure through PSI, swap, compressed memory, cgroup isolation, and controlled responsiveness experiments.
Broader goal
How can memory-management policy preserve responsiveness when RAM is constrained?
Evidence to date
Controlled reruns show PSI before material swap growth, and a corrected canary records pressure-related responsiveness degradation. A three-host V5 study found that cgroup-memory growth rate identifies the tested moderate pressure regime about five seconds before PSI-only qualification and falls back cleanly under gradual pressure; earlier activation did not consistently outperform static zram.
V1–V5 established early PSI/swap ordering, a valid responsiveness measure, compression sensitivity, and a reliable rate classifier with an inconclusive performance result. Research is ongoing. The evidence does not establish one globally beneficial guided-zram policy: static zram was best in the focused moderate/structured comparison, gradual/random outcomes were host-dependent, and activation timing alone is insufficient to explain responsiveness.
Earlier work in Nepali legal information retrieval, embedding evaluation, and AI-assisted research workflows.
A controlled ablation of bounded reclaim for PSI-guided compressed memory under constrained physical memory.
Study contribution: Evaluates pressure-guided memory-management policies across foreground workloads, with explicit attention to responsiveness and workload-dependent effects.
An empirical study of embedding choices for retrieving Nepali legal documents in a low-resource setting.
Study contribution: Compares retrieval behavior rather than treating embedding choice as an implementation detail, with attention to the constraints of Nepali legal text.
A lightweight pipeline that generates search terms, retrieves open-access papers, ranks them with transformer embeddings, and applies statistical thresholding.
Study contribution: Evaluates three embedding models and documents a practical preliminary-review workflow without claiming ground-truth relevance labels that were not collected.
A retrieval-augmented legal analysis pipeline designed for access to information in the Nepalese legal context.
Study contribution: Processes legal documents for retrieval, expands user queries, and evaluates rule-recall, rhetorical-understanding, and interpretation-oriented questions.
Selected Rust systems projects involving execution state, persistence, workspace restoration, and terminal interfaces.
A Rust notebook interface for exploratory Python work, using a human-editable Markdown format with Jupyter import/export and checkpointed runtime state.
Systems focus: Execution state, notebook serialization, kernel management, environment selection, and terminal interaction.
A Rust terminal multiplexer centered on declarative project layouts, persistent sessions, and best-effort workspace restoration. It grew from an earlier tmux session helper, Smart Session Manager.
Systems focus: PTY management, session persistence, and the boundary between shareable project configuration and machine-local runtime state.
Selected experiment logs, methodology notes, and reading notes from ongoing systems work.
What PSI/swap validation established, why memory-growth rate classifies the tested ramps early, and why earlier zram activation is not yet a general responsiveness win.
Docker packages layered filesystems with metadata and uses Linux namespaces, cgroups, and the shared host kernel to run isolated, reproducible environments without full virtual machines.