Daniel S. E. Hill

Broker-agnostic trade execution layer: signal validation chain, pre-trade risk gateway with registerable checks, order/fill/position lifecycle, a seeded simulated broker, and a three-state outcome-level circuit breaker with drawdown-scaled position sizing. Contains no strategy by construction — signals arrive as injected objects.

LLM provider failover: same model, same history, next provider when one fails. Per-gateway circuit breakers with exponential cooldown, an error taxonomy separating gateway-level from model-level failures, SSE stream-restart semantics, and an OpenAI-compatible proxy. Adding a provider is one YAML entry.

Hot/warm/cold object store with per-payload codec selection and a FastAPI control plane. Benchmarked with the causes written up — including a 288× store/retrieve asymmetry traced to triple compression in auto mode. The Rust core ships as source, documented as never compiled; no claim is made for it.

A Tkinter desktop app and a dependency-free offline-first PWA writing one dataset with no server between them. Devices converge through per-item last-write-wins merging and deletion tombstones; the merge is order-independent and idempotent, and the tests prove it.

Dependency-free AST-based cyclomatic complexity analyser: per-function scoring, risk banding, JSON/Markdown reports, and trend tracking so complexity drift is visible across runs. Documents exactly where it deviates from strict McCabe — and ships an analysis of its own source as the worked example.

Data standardization plus a four-tier graceful-degradation ladder — retry, fallback calculation, simplified operation, emergency fallback — with domain-specific fallbacks per operation type, quality scoring, and an IQR outlier budget that rejects bad data rather than silently repairing it.

Range-based volatility estimators — Parkinson, Garman-Klass, Rogers-Satchell, Yang-Zhang — side by side so the numbers can disagree visibly, plus adaptive Bollinger/ATR/Keltner analysis and diagnostics for when GARCH-family models are, and are not, appropriate. A 215-method suite.

Advanced statistics for noisy series: L-moments, robust statistics, bootstrap analysis, extreme-value analysis, nonparametric tests, and a Bayesian block — conjugate posteriors, change-point detection, Bayes-factor model comparison, and posterior predictive checks.

An autonomous mathematics research engine: seven pluggable research modules — Collatz, Goldbach, prime distribution, graph theory, symbolic regression, cellular automata, dynamical systems — under expected-value task scheduling with psutil-enforced compute budgets, and an independent verification engine that re-derives findings rather than trusting the module that produced them. Ran unattended for 3 days / 107 research cycles.

A binary-lattice physics laboratory: measured emergent inverse-square attraction (exponent 2.09), mass proportionality r = 0.891, FFT-accelerated long-range solves, a ring-decomposed retarded potential (causality at one cell per tick, static-grid equivalence to 1e-14), and a THEORY_CONFLICTS.md recording where the results contradict the theory.

From-scratch two-state Gaussian HMM, a strictly walk-forward regime gate with bit-exact no-lookahead tests, and a Politis–Romano stationary bootstrap for Sharpe confidence intervals. Ships the negative result it was built to find: every gated variant underperformed the ungated strategy (+0.33 vs +0.21/+0.20/+0.24).