More Projects¶
A selection of smaller tools, experiments, and unfinished research prototypes. Each entry reflects the project's documented state.
Trident-TS¶
Architecture implemented · Not trained
Trident-TS is an experimental neural architecture for weather forecasting.
I designed a Vision–Memory–Decoder architecture combining DINOv2 visual features, an Echo State Network, a Mamba state-space model for temporal processing, and a quantile prediction head producing P25, P50, and P75 forecasts.
The data pipeline uses WeatherBench 2 with Xarray, Zarr, Dask, and Google Cloud Storage.
The architecture and data-loading infrastructure were implemented, but the model was not trained or evaluated. No forecasting accuracy or inference speedup was demonstrated.
Alpaca¶
In progress
Alpaca is a command-line interface I developed to simplify model management and serving with vLLM.
Inspired by Ollama's model-management workflow, it provides commands for downloading models, starting inference servers, managing running processes, and configuring inference parameters.
I also implemented commands for Ray cluster management and distributed serving. The repository documents cluster initialization, worker registration, monitoring, and shutdown.
Local model serving is functional. The distributed workflow has not been established as production-ready or validated through a documented multinode benchmark.
Hugging-Hat¶
Infrastructure implemented · Not trained
Hugging-Hat explores adding optional computation modules to a frozen decoder-only language model while preserving the standard Hugging Face generation interface.
I implemented three components:
- Thinker Hat: a parameter-shared recurrent module for additional computation.
- Latent Router Hat: selects a computation budget from a predefined set of iteration counts.
- Critic Hat: uses cross-attention over cached prompt representations.
The implementation includes forward-hook integration, configuration files, checkpoint serialization, and tests for module attachment and persistence.
The infrastructure was implemented and tested, but the modules were not trained. No improvement in language-model performance was demonstrated.
Toxic Comment BERT PEFT¶
Completed
I developed a training and serving pipeline for toxic comment classification using BERT Base Uncased and LoRA.
The project uses Hugging Face Accelerate for training on balanced samples from Civil Comments. A FastAPI service exposes classification predictions and confidence scores.
I also implemented Docker packaging, ONNX export support, and scripts for deployment and monitoring through AWS SageMaker and CloudWatch.
Training accuracy and F1 were tracked, but no final evaluation scores are documented. The project primarily demonstrates the integration of parameter-efficient fine-tuning with a model-serving workflow.
Blackhole¶
Completed
Blackhole is a Python command-line utility I developed to simplify one-way file transfers to remote machines.
It wraps rsync over SSH, providing a shorter interface for transferring files and directories while retaining resumable transfers, SSH key authentication, and optional destination paths.
The tool delegates the actual file transfer to rsync rather than implementing a new transport mechanism.
Telecommit¶
Completed
Telecommit is a Python CLI that generates Git commit messages from repository changes using the Anthropic Claude API.
It reads staged changes, optionally includes unstaged changes, and generates a suggested commit message. An additional option allows the tool to execute the commit.
Configuration is managed through a YAML file and environment variables.
Because the tool sends Git diffs to an external API, it is intended for repositories where sharing those changes is acceptable.