Technical publication
Labs
Reproducible experiments and benchmarks, with explicit environments, methods, results, failure cases, and limitations.
Engineering claims become useful when you can examine the method and reproduce the result. These labs explore inference, retrieval, quantization, and agent execution through bounded experiments.
What a lab should tell you
Look for the question, software and hardware versions, workload, measurement method, observed results, and failure cases. Treat a result as evidence for that environment; check the limitations before applying it elsewhere.
Start with these experiments
- Benchmark LLM Serving Without Lying to Yourself
- Serve a 3B LLM on CPU With llama.cpp
- When Q4 Passes Latency but Fails Quality
- Build a Lexical Search Baseline Before Adding Embeddings
- Build a Fully Offline RAG App in Your Browser
These links point to existing experiments. Newly classified lab reports will also appear in the recent-articles list below. Browse Writing for the complete archive.
Latest
Recent articles
No articles have been published in this section yet.