Liam Grinstead

What I can offer and help with: I build transparent, reproducible agent systems that solve hard “state + decision” problems where standard ML pipelines stall, drift, or become unreproducible. My work focuses on decision-timing under uncertainty, durable state management, and turning every run into an inspectable artifact with cryptographic lineage. If you’re working on agents, automation, or research demos and you keep hitting the same wall—“why did it do that, and can we reproduce it?”—this is exactly what I build. I can help with • Agent state durability: reproducible memory/state handling across retries, branching, tool calls, and multi-agent handoffs (planner/executor/reviewer) without mystery behavior. • Decision-timing frameworks (beyond standard ML): systems that act when the cost of waiting exceeds the cost of acting—explicit commit/collapse logic, failure modes, and audit trails. • Non-standard programming approaches: collapsing complex behaviors into simpler, verifiable primitives (thresholds, feedback loops, cascades) instead of brittle, overfit heuristics. • Symbolic agents & “entangled” influence models: reflex/instinct/reflective/meta agents with explicit coupling rules that can be inspected and stress-tested. • Reproducible artifact lineage: every run becomes a “Codex” record (inputs → intermediates → decisions → outputs) sealed with hashes so results can be verified later. • High-performance simulation + benchmarking: fast NumPy/Numba-style simulation work where performance metrics are measured and reported as part of the experiment. • AI × quantum / computing research prototyping: practical, testable toy-models that connect agent collapse dynamics to computation constraints (latency, throughput, scaling), without hand-wavy claims. Core focus Rendered Frame Theory (RFT): a collapse/decision framework I developed to model complex adaptive systems using thresholds, feedback loops, and cascade dynamics—designed for open inspection and reproducibility. Best-fit collaborations • People building multi-step agent pipelines who need reproducibility and explainability. • Researchers shipping demos who want falsifiable runs and durable logs. • Builders optimizing performance and looking for clean simulation kernels + measurable benchmarks. • Anyone tired of black-box “agent magic” and wants explicit rules, explicit data, explicit failure modes. Everything I publish is designed to be inspected, reproduced, and argued with.