Industry insights, documented case studies, and analysis from the White Oak engineering team. When we publish, it means something is worth your time.
Every post here represents work we've done, a perspective we've earned, or a result we can stand behind. The Intelligence Log includes both technical insights from our practice and anonymized case studies written to answer one question: does this approach apply to my problem?
The loop strategy raises survival probability from \(2^{-100}\) to 31% — not by improving any individual prisoner's odds, but by correlating all 100 outcomes through a single permutation cycle structure. A complete proof via harmonic numbers, the exact failure sum \(\sum_{L=51}^{100} 1/L \approx \ln 2\), and a 100,000-trial Python simulation showing zero random wins versus 31,182 loop wins.
Should you switch doors? Yes — always. Exhaustive case enumeration, a complete Bayesian derivation showing \(P(D_2 \mid H_3) = 2/3\), the generalized N-door result, and a 1,000,000-trial Python simulation that leaves no room for doubt. Plus why this is a Bayesian updating problem in disguise.
The optimal strategy in the Do-Over Game — redraw or keep? — turns out to be the golden ratio. A full Nash equilibrium derivation showing how \((t^*)^2 + t^* - 1 = 0\) produces \(t^* = (\sqrt{5}-1)/2 \approx 0.618\), with Python simulation and win-probability visualization.
A classic Goldman Sachs / Morgan Stanley quant interview question. Starting with one amoeba that can die, stay, or split into two with equal probability, is extinction certain? Yes — and the proof via the Law of Total Probability and a quadratic fixed-point equation takes exactly three lines.
You flip a fair coin. Expected flips to see HH: 6. Expected flips to see HTH: 10. Both answers surprise almost everyone. A complete derivation using state-transition equations, recursive algebra, and a 100,000-simulation Python proof — plus why overlapping pattern structure drives the difference.
Most middle market valuations are built on a single set of assumptions. Monte Carlo simulation produces a probability distribution across 10,000 scenarios instead. Here's the math, the Python implementation, and why the difference matters when capital is on the line.
Algorithmic trading infrastructure has zero tolerance for data loss, latency spikes, or silent failures. This is the four-layer architecture — WebSocket ingestion, ring buffers, Z-score anomaly detection, and circuit breakers — that survives production.
Apps Script time-based triggers can automate daily data exports, weekly executive reports, and multi-system syncs without a single SaaS subscription. Here is how to build a production-grade automation layer for your Google Workspace operation.
Off-the-shelf CRMs were designed for the median B2B sale. When your deal flow is highly specialized — equipment financing, niche PE, distressed assets — you need a custom build with AI intelligence at the core. Here is the architecture and implementation.
Manual client reporting is a 15-hour-per-week problem disguised as a communication process. The fix is an ETL pipeline — Python, SQL, Google Sheets API — that generates polished, accurate reports on a schedule without human involvement.
Financial services sites are YMYL — Google holds them to its strictest quality standard. Here's the technical foundation that earns durable rankings: E-E-A-T signals, JSON-LD schema markup, Core Web Vitals optimization, and topic cluster architecture.
MAPE alone misleads — it breaks near zero, penalizes overforecasting asymmetrically, and carries no baseline comparison. The full evaluation stack: MAPE, RMSE, MASE, Theil's U, and Ljung-Box residual autocorrelation testing, with Python implementation.
You don't need Snowflake, dbt, or a dedicated data engineering team to get live KPI tracking. A watermark-based incremental query layer and a stateful compute engine can power a real-time operations dashboard from your existing PostgreSQL database in under a week.
Most RAG systems fail in production not because the LLM is wrong, but because the retrieval layer is broken. Here's the full architecture — pgvector, chunking strategy, idempotent indexing, and retrieval evaluation — built for financial document search.
A company can be EBITDA-positive and still default on debt. Standard P&L analysis doesn't reveal this risk — a cash flow waterfall model does. Here's the Python implementation with DSCR output, tranche prioritization, and stress testing.
A regional operator was losing $8,000–$10,000 every month with no identifiable cause. Twelve months of POS data, a Monte Carlo simulation, and a shift-level heatmap revealed exactly where — and why. The result: an $18,000 monthly swing in under 90 days.
Starting from page 27 with zero automated leads, a complete technical SEO rebuild and lead-capture architecture produced a 58% top-of-SERP rate and turned organic search into the firm's primary growth engine — attributing $67.7M in new business volume.
In a high-velocity data environment, human monitoring latency was costing capital on every shift. A C# and Python system using WebSocket ingestion, Z-score normalization, and adaptive regression eliminated the human-in-the-loop — detecting anomalies in under one second across 10,000+ data points.
An industrial e-commerce operation was running entirely on spreadsheets, email chains, and disconnected platforms — with zero executive visibility into the pipeline. A purpose-built proprietary CRM centralized every workflow into a single real-time interface with no SaaS subscription required.
Standard software cannot handle distressed debt portfolios where account histories are incomplete, interest compounds irregularly, and the output must be court-defensible. A dual-engine architecture — path-dependent compounding and regression-based appraisal — produced auditable results that held under legal scrutiny.
A financial advisory firm's decades of proprietary documentation were functionally inaccessible — buried in directories, queryable only by those who already knew where to look. A Retrieval-Augmented Generation system made the entire knowledge base answerable in plain language, in under two seconds, with cited sources.
Every case study on this page started with a conversation. If one of these problems sounds familiar, that's a good reason to reach out.
Initiate Contact