Projects

Open-source benchmarks and tools. Each one is reproducible from its repository, and most have a write-up on the blog.

RAG 2026 Benchmark

Nine retrieval strategies - hybrid search with and without a reranker, MetaRAG, agentic RAG, A-RAG, Search-R1, LightRAG and Graph-R1 - run on the same OpenSearch index, generator, judge and 1,000 questions across five datasets.

Turning on the generator's reasoning moved accuracy more than any retrieval architecture.

Python · OpenSearch · LightRAG · Graph-R1 · Search-R1

Read the write-up

LLM Inference Benchmarks

vLLM, SGLang and TensorRT-LLM serving Qwen3-32B on a single NVIDIA L40S, measured with AIPerf across concurrency scaling, long context, FP8 / AWQ / GPTQ quantization and a MoE model.

A mid-range GPU most published benchmarks skip in favour of the H100.

Python · vLLM · SGLang · TensorRT-LLM · AIPerf

Read the write-up

Multi-Model Serving on Triton

Three ONNX vision models served from one GPU with NVIDIA Triton Inference Server, with deployment notebooks for AWS SageMaker, Azure ML and GCP Vertex AI.

Up to 90% lower GPU serving cost than one instance per model.

Triton · ONNX · SageMaker · Azure ML · Vertex AI

Read the write-up

AI Research Radar

A Python pipeline that discovers AI papers on arXiv, enriches them with citation and code metadata, ranks them deterministically and drafts a weekly radar of the papers worth reading - with a separate critic reviewing every analysis.

Python · arXiv API · SQLite · OpenAlex · LLMs