Infrastructure as Code for Reproducible Science

This repository automates the provisioning of standardized research environments, ensuring that local development (macOS) and remote training (HPC) share strictly consistent dependencies. It integrates:

1. Local NLP: An offline RAG pipeline (Gemma 2) for private, semantic search over research literature.

2. Writing Automation: A continuous LaTeX build system for real-time manuscript validation.

3. Rapid Reference: Instant fuzzy retrieval for opening canonical textbooks, citations, and knowledge graphs (Obsidian).

Designed to minimize technical friction and guarantee experimental reproducibility.

The repository is structured to separate package declaration from installation logic, ensuring modularity and easier auditing of system dependencies.

├── manifests/          # Declarative lists of system packages (Brew, Python, Obsidian)
├── llm-config/         # Local LLM model weights and version-controlled configurations
├── scripts/            # Idempotent automation logic
│   ├── install_file.sh # Module-level installer
│   ├── install_folder.sh # Batch provisioning utility
│   └── local_rag/      # Custom pipeline for offline literature retrieval
├── configs/            # Application-specific configurations (Tmux, Alacritty)
└── keyboard/           # Low-level input remapping

Objective: Eliminate the "works on my machine" class of failures that plague collaborative research. By replacing imperative setup commands with declarative manifest files (manifests/), this system guarantees that the local development environment and remote training nodes are mathematically identical. This prevents environmental drift—bugs caused by subtle version mismatches in CUDA, compilers, or Python libraries—ensuring that experimental results are strictly reproducible across heterogeneous hardware.

Objective: Enable the analysis of sensitive, embargoed, or clinical data (e.g., MIMIC-IV) without compliance risks. The environment provisions a completely air-gapped inference stack (Gemma 2 via Ollama) and a local RAG pipeline. This architecture ensures zero data egress, allowing for the use of powerful LLM assistants on confidential datasets where cloud-based APIs (OpenAI/Anthropic) are legally or ethically prohibited.

Objective: Mitigate the productivity loss caused by unstable connections to High-Performance Computing (HPC) clusters. Optimized for headless operation on supercomputers (e.g., NYU Greene), this workflow employs session persistence (via Tmux) to decouple the user interface from the execution state. This ensures that long-running training jobs and unsaved code changes survive network interruptions or VPN disconnects without process termination.

Utility: Accelerates literature review by enabling conversational interrogation of dense technical papers without internet dependency. Impact: Reduces the time required to synthesize findings from large collections of PDFs. Implementation: A custom pipeline (scripts/local_rag) leverages ChromaDB and SentenceTransformers to provide low-latency, semantic search over local knowledge bases.

Utility: Decouples paper writing from cloud-based editors (Overleaf), removing reliance on internet connectivity and third-party servers. Impact: Provides a "live editor" experience with instant feedback, while enforcing version control on the source text. Implementation: A continuous build wrapper (scripts/compile_latex) uses filesystem monitoring to trigger incremental compilation and manages artifact hygiene to keep repositories clean.

Utility: Drastically reduces the time spent screening literature for specific formulas, citations, or keywords. Impact: Enables rapid verification of prior art across gigabytes of unindexed PDFs where standard system tools are too slow or imprecise. Implementation: A multi-threaded keyword spotter (scripts/search_pdfs_no_llm) that outperforms grep by parallelizing text extraction and generating structured HTML reports for review.

Utility: Minimizes the friction of accessing standard reference materials during deep work sessions. Impact: Eliminates the cognitive load of navigating file hierarchies, allowing for instantaneous context switching to textbooks. Implementation: A CLI wrapper (scripts/search_textbooks_by_name) that uses fuzzy string matching to resolve partial mental queries (e.g., "deep learn") into immediate file access.