Notes & Writing

Thoughts on compiler design, inference plumbing, and why determinism beats heuristics.

The Flattening Tax: Why LLMs Shouldn't Read Raw JSON

Modern frontier models spend 60% to 80% of prompt tokens on serialization boilerplate. The case for compiling structured data into a canonical IR before inference.

Determinism Over Heuristics: Why Prefix Caching Demands Canonical Serialization

Note

KV-cache hit rates depend on bit-for-bit stability. Why the context engineering layer beneath the model must be strictly deterministic.

Inference-Native Memory: Beyond Approximate Vector Retrieval

Spec

Vector similarity search is insufficient for structured analytical context. Why autonomous systems need typed, content-addressed memory substrates.

The Token Cost of Syntax: An Empirical Audit of BPE Tokenization

Benchmark

Empirical measurements of how standard Byte-Pair Encoding splits structured records, and where token bandwidth is actually lost in production pipelines.