Case studies

Each demo on this site compresses a production pattern into something you can poke at in a browser. These are the longer stories: the problem each pattern exists to solve, how the method works, where the real version runs, and what it changed.

Telemetry Anomaly Lab

Problem: high-consequence systems fail quietly — drift, stuck sensors, and dropouts hide inside normal-looking dashboards until they are expensive.

Approach: run several cheap online detectors in parallel (adaptive z-score, EWMA, CUSUM, variance-collapse), each blind to different failure modes, and treat their disagreement as the signal to triage. Baselines learn only from calm samples so a fault cannot poison its own detector.

In production: the same pattern watches this site's own infrastructure — service health fans, backup freshness with a 26-hour dead-man threshold, and state-change-only alerting so a quiet pager still means something.

Result: faults surface in seconds on the demo, and the production watchtower has caught real regressions (a silently failing semantic-search endpoint, a stale backup) before any human noticed.

Entity Resolution Workbench

Problem: enterprises pay twice, report wrong, and mis-count because 'Northwind Indust. Supply Inc' and 'NORTHWIND INDUSTRIAL' are one vendor the database thinks are two.

Approach: blocking to make comparison tractable at scale, Jaro-Winkler similarity across weighted fields with missing-field weight redistribution, then transitive union-find clustering — with the match threshold exposed because it is the real business decision.

In production: the same discipline underpins two decades of supplier financial management across 50+ suppliers and enterprise reporting where clean identity was the difference between reports leadership trusted and reports they audited.

Result: the demo deduplicates twelve damaged records live in the browser; the transferable piece is knowing where chain-merge failure lives and designing the threshold review around it.

Family Whole-Life System

Problem: family finances live in disconnected apps, so nobody can answer the compound questions: are we saving enough, are we protected, what breaks first?

Approach: one shared state, four derivations. A double-entry ledger with a books-balance invariant feeds actual spending into a 400-path Monte Carlo retirement engine, DIME insurance adequacy, an umbrella-liability trigger, and a scored estate checklist — so editing one categorization rule reprices the entire plan.

In production: the real system runs privately on Cloudflare Workers and D1 for one household, with the same invariants: balanced journals, paired transfers, evidence-graded entries.

Result: a financial picture that audits itself — the demo's deliberate gaps (underinsured home, missing beneficiary) show the system finding problems, which is its job.

Workout Tracker

Problem: training plans die because week-to-week progress is invisible and overreach is only obvious in hindsight.

Approach: one honest table of sets; everything else is derivation — Epley e1RM trends per lift, weekly tonnage, and chronological PR detection that survives a deload without panicking.

In production: the server-side version keeps history in D1 across devices; the demo keeps user-entered sets in the browser only.

Result: measure-derive-alert on a barbell — the same loop that runs sales pipelines and server fleets, small enough to read in one sitting.

Manuscript Studio

Problem: in the AI era the hard questions about any text are 'is this exactly what the author wrote?' and 'can the claims be checked?'

Approach: content-address every paragraph with SHA-256 so edits are detected and must be accepted as explicit revisions; machine-extract typed claims (quantitative, attributed, normative) into a human verification queue; derive the concept map from the text itself.

In production: the same architecture backs a book-length manuscript served by a knowledge API that can prove it quotes the manuscript verbatim.

Result: prose an automated assistant can responsibly serve — hash-verified, claim-audited, revision-explicit.