work / selected systems
things i’ve built
and measured.
llm inference, production ml, multimodal research, and evaluation systems—with the constraints and trade-offs left visible.
latency-aware speculative decoding
A correct speculative decoding engine, benchmark harness, and GPU profiling study for Qwen language models.
verifiable preference optimization for reasoning llms
A verified reasoning-data pipeline and LoRA DPO study with careful validation of truncation and answer parsing.
temporal knowledge-graph memory
Graphiti and Neo4j memory for clinical LLM agents, evaluated with a six-stage verification pipeline.
structure-grounded medical qa
RDF and SPARQL retrieval with claim-level verification for faithful medical question answering.
edi-summ: multimodal clinical summarization
Image-guided BART with decoder cross-attention for multimodal clinical document summarization.
joomla 4 guided tour
An MVC component and plugin shipped into Joomla 4 core through Google Summer of Code.
unifyai / ivy numerical core
Cross-backend numerical primitives for PaddlePaddle and PyTorch, including a BLAS/LAPACK-backed dot implementation.