Long-form technical writing. Each post starts from a real decision record (context, options considered, what I chose, what it cost) and expands it into the reasoning I'd want to read if I were checking someone else's work.
Synapse's knowledge graph moved from an in-memory NetworkX object to a SQLite-backed schema. What moved, what stayed compatible on purpose, and the benchmark I still owe.
Synapse routes across a wide provider list behind one router. A rate limit, an auth failure, and a context overflow all need a different response, and the honest cost of how it tells them apart.
Four channel adapters, one shared ingestion layer. What FloodGate, dedup, and a capped asyncio queue guarantee under a concurrent burst, and what the evidence for it actually is.
tldr skips a model call wherever static analysis can answer instead. The three-tier pipeline, the failure that reshaped its token budget, and which of its numbers are still unverified.
I ran Qwen3-Coder 30B on an 8 GB RTX 3060 Ti, hit an OOM, pivoted to TurboQuant, applied a CUDA DLL fix, and reached 32.49 tokens/sec on a 262,144-token context with about 500 MB of VRAM to spare.
Feed: RSS. Append .md to any post's URL for the plain-text version.