receiving · processing · transmitting
i make language models reliable in production — evals that catch real failure modes, agent pipelines that ship, open-source tools that cut the bullshit out of llm systems.
dendrites take the world in. axons fire only what matters.
research, patterns, techniques arrive as raw signal. i compress them into systems that work under real load — not demo conditions, not vibes, not cargo-cult evals.
currently building at sidecar. on the side: tools for people who have to trust an llm with something that can break.
hallucination detection — semantic entropy, logprobs, web grounding, llm judge. confidence you can ship.
llm observability — tracing, behavior eval, security scanning, alerts. see the system before it fails loud.
context that earns its tokens — 85% reduction, 99.2% quality retained in evals.
semantic double-pass chunking for rag. boundaries by meaning, not character cosplay.