Implementations.

Open source implementation and extension of Google Research's PaperBanana for automated academic figures, diagrams, and research visuals, expanded to new domains including slide generation.

A framework for building reliable LLM-based reasoning systems that adaptively allocate computational resources under budget constraints. Combines adaptive search strategies, uncertainty estimation, and verification cascades to maximize reliability while respecting token, tool-call, and verification budgets.

KV-cache compression for LLMs. Reference implementations of TurboAngle and TurboQuant codecs with Triton GPU kernels for inference time KV cache reduction.

Fast LLM inference with 2.8x speedup using speculative decoding. A streaming implementation focused on practical end-to-end latency improvements.

Implementation and evaluation of the Scaling Embedding Layers in Language Models research paper.

Implementation of ReaRAG, a knowledge-guided reasoning model that enhances factual accuracy using iterative retrieval-augmented generation. Adapted for modular LLM integration and customizable retrieval strategies.

Implementation of the Better Embeddings with Coupled Adam research paper.

Implementation of the research paper DarwinLM: Evolutionary Structured Pruning of Large Language Models.

ParamΔ (Parameter Delta) implementation. Enabling continuous model post-training at no training cost through parameter delta merging.