---
title: "Work — Platforms in production | Jake Williams"
canonical: https://jakeawilliams.com/work/
description: "Five bodies of platform work plus the open-source proof: the Claude plugin marketplace (12 plugins, 6 domains, ~200 monthly users), semantic analytics in production, the CDC platform rescue, observability as code, the enterprise data architecture foundation, Waybill, and Lobby Ledger, a public lobbying explorer on a Databricks lakehouse."
---

# Platforms in production.
Five bodies of platform work — plus the open-source proof — written the way it actually happens: constraints first, numbers where they're real, diagrams instead of proprietary detail. Together they are one system — the architecture set the foundation, the streams feed the lakehouse, the lakehouse feeds the semantic layer, and the whole thing is observed as code.

01 · Flagship · AI platform
## The Claude plugin marketplace
Created, architected, and run: an enterprise marketplace of 12 Claude plugins with authored domain skills spanning six business domains, enabled company-wide as the semantic-analytics solution. ~200 monthly active users — 40% of the company, five C-suite regulars — run 2K+ queries a month at 30% lower token consumption than unassisted LLM use, assembling QBRs, reserve reviews, financial reports, and canonical dashboards. I own the roadmap, the model evaluations, the governance, and the support.
Read the case study →
Plugins 12
Business domains 6
Monthly users ~200 · 40% of company

02 · AI & semantic analytics
## Semantic analytics in production
A production Claude + Snowflake Cortex layer giving Claims, Policy & Underwriting, Finance, and Legal natural-language analytics over governed data. The quarterly reserve review went from 3 days to under 30 minutes, QBR prep from 4 days to under an hour — with consistent regulatory reporting and cross-domain metrics (policy erosion among them) that were previously impractical.
Read the case study →
Reserve review 3 days → <30 min
Domains in production 4
Stack Claude · Cortex · dbt

03 · Streaming platforms
## The CDC platform rescue
Inherited a half-built replication platform mid-build, shipped it end to end, then re-architected it: Azure Event Hubs → Apache Kafka (Strimzi on Kubernetes) — Debezium CDC connectors across SQL Server, PostgreSQL, and Oracle (LogMiner), plus custom Java source connectors where none existed. The event-driven backbone now feeds every downstream domain, with type-2 period-fact modeling enabling historical questions, Dagster sensors alerting policy teams inside their decision window, platform-wide GitOps delivery via Argo CD and Helm, and PySpark compaction jobs that keep Apache Iceberg tables query-efficient as CDC writes accumulate.

Platform cost cut $30K+/yr
Ops Helm · ArgoCD · GitOps

04 · Observability
## Observability as code
Zero dashboards built in a UI: Datadog and Prometheus dashboards and monitors ship exclusively through GitHub Actions CI/CD. Forecast monitors predict OOM and strain — 5 OOM incidents prevented and ~20% infra cost avoided by scaling on signal instead of overprovisioning. Observability run as FinOps.

As code 60+ dashboards/monitors
Scaling on signal, not fear

05 · Enterprise architecture
## The data architecture foundation
Before the streams and the semantic layer, the foundation: served as reporting architect on the core design committee that designed the enterprise data architecture end to end — data lake, EDW, and the reporting layer — and helped build the reporting solution the company scaled on, with core reports for Underwriting, Finance, Claims, and Operations. I run the BI tenant today, so the governed number is the same in every surface that shows it.

Scope lake · EDW · reporting
Stewardship BI tenant, run today
Outcome scaled with the company

06 · Open source
## Waybill — bring receipts
An open-source Claude Code plugin for token accounting on AI-assisted work: it deterministically attributes agent token spend to the work items that shipped — metered from local transcripts and git history, never estimated — with evidence tiers, sealed pre-registered estimates, per-session conservation checks, and a verification pack the recipient can re-run offline. Public code, public docs, CI on every commit: the quality bar is inspectable.
github.com/Jakeintech/waybill ↗
Attribution metered, not estimated
Integrity verifiable offline
License MIT · plugin marketplace

07 · Open source · Lakehouse
## Lobby Ledger
A public explorer of U.S. federal lobbying disclosures: who lobbies Congress, for whom, on what bills and issues, for how much, quarter by quarter. Every registrant-client engagement that reported a $500k filing in 2024 or 2025, with its full history, flows through a complete Databricks lakehouse in one public repo. Bronze lands the Senate's filings as they arrived. Silver is a Data Vault 2.0 Raw Vault: hubs, links and satellites keyed by md5 hash keys, every insert guarded by hash_diff so a rerun inserts nothing. Gold is one mart per page, with the arguable rules (which amendment wins, what counts as income, how a filing's income is attributed to its issue codes) stated once in SQL and tested. On top, a React app with topic, bill, client, firm and lobbyist pages, each built around a quarterly chart, reading Gold through a single read-only identity. The SQL, the tests, the proof queries and the decision records are all in the repo.
Open the live app ↗ github.com/Jakeintech/lobbying-experience-search ↗
Entity pages 5
Rerun duplicates 0
Stack Databricks · Data Vault 2.0 · Unity Catalog · AppKit

Want the detail behind any of these — architecture, tradeoffs, costs? That's a good first conversation.
Email info@jakeawilliams.com