Turning on the generator's reasoning moved accuracy more than any retrieval architecture.
Python · OpenSearch · LightRAG · Graph-R1 · Search-R1
Read the write-upMichał Wojdylak
AI Infrastructure Engineer
AI Lead at WPP Enterprise Solutions · Poznań, Poland
Building production AI systems, LLM infrastructure, inference platforms and cloud-native ML solutions.
Selected projects
All projects →A mid-range GPU most published benchmarks skip in favour of the H100.
Python · vLLM · SGLang · TensorRT-LLM · AIPerf
Read the write-upThree cloud multi-model patterns compared on what each makes easy - not a cost benchmark.
Triton · ONNX · SageMaker · Azure ML · Vertex AI
Read the write-upLatest writing
All posts →
9 min read
KV Cache: The Hidden Variable in LLM Inference Benchmarks
How much KV cache an inference engine reserves decides how many requests it can run at once, and that often matters more than the engine. Measured on Qwen3-32B with vLLM, SGLang and TensorRT-LLM on a single L40S, with the flags to control it in each engine.
14 min read
Agentic RAG, Graph-R1, MetaRAG and Friends on One Search Engine: What Actually Helps in 2026?
Nine retrieval strategies from 2024-2026 - naive and reranked hybrid search, MetaRAG, agentic RAG, A-RAG, Search-R1, LightRAG and Graph-R1 - benchmarked on the same OpenSearch index, the same base generator and the same 1,000 questions. In this benchmark the biggest lever turned out to be the generator's reasoning, not the retrieval architecture.
20 min read
LLM Inference Benchmarks: vLLM vs SGLang vs TensorRT-LLM on a Single L40S
A systematic comparison of three leading LLM inference engines - vLLM, SGLang, and TensorRT-LLM - benchmarked on Qwen3-32B across concurrency scaling, long-context, quantization, and MoE workloads on a single NVIDIA L40S GPU.