Hi, Iām Kit Kyo ā a full-stack engineer who works across the whole stack: product code, DevOps, and agent development (building LLM-driven agents and the plumbing around them).
Right now Iām putting serious time into machine learning and infrastructure ā going from āship the productā to understanding the systems underneath. This blog is where I keep the notes, experiments, and write-ups from that journey.
Recent posts below ā or read more about me.
Kit Kyo
Cheap factors, four domains: a scorecard ā and why 'no single factor works' isn't 'no signal'
We put interpretable, locally-computable factors from four mechanistically-different domains ā software-defect prediction, paper acceptance, AI-text detection, and a world-model physics proxy ā on one shared per-factor scorecard. Predictive power has a domain-dependent ceiling, the factors donāt transfer across domains, and a single-factor reading nearly had us call an industrial benchmark a ācollapseā when the signal was just multivariate.
Does a multi-agent panel beat a single LLM at evolutionary code search? (We tested it. No.)
We replaced the single-LLM mutation step of an AlphaEvolve-style evolve loop (OpenEvolve) with a proposerācriticāaggregator panel and tested it rigorously across three problems. It never won ā equal on easy optimization, significantly worse on hard optimization (p=0.01), and on correctness-gated ARC it cracked nothing a single call couldnāt. A clean negative result, with the baseline/ablation harness that makes it trustworthy.
The moat in LLM chip design isn't the model ā it's the environment
A field note on 2025ā2026 agentic RTL generation: every result that moved the needle came from the verification loop wrapped around the LLM, not the LLM itself. For someone who builds compilers and ML infra, that loop is the product.
Kill the tail, not the cores: profiling the verifier under RL-with-verifiers
As RL-with-verifiers scales, the Lean verifier is both the reward source and the throughput bottleneck ā a CPU/scheduling problem almost nobody benchmarks on commodity, on-prem hardware. We profiled it: on an 8-core box, ~4.5% of proofs run away and eat most of the single-thread wall, and a tuned per-proof deadline beats adding cores. The honest-measurer story (including the claim we had to retract) is the point.
Your edge LLM can't tell time ā and a 40-line memory can fix it (for some models)
A tiny adversarial benchmark shows small local LLMs confuse token-distance for elapsed wall-clock time. A zero-dependency continuous-time memory fixes it for sub-millisecond cost ā but whether a model will even accept that fix turns out to be a model-dependent deployability property that capability leaderboards donāt measure.