Rick Huang

13 published · 3 in the works

Tech Talk

Data architecture and modeling across industries, and AI that does real work inside the data stack.

  1. 15 min readGreen pipeline, wrong numbers: data quality at every layerA pipeline that ran is not a pipeline that is right. Part 2 of a series on data foundations for enterprise AI: the checks that run at every layer, the SLA on the gold layer, and who gets told when something breaks.Data Architecture
  2. 8 min readBefore you build the agent: working backwards to your data pipelineAn agent is only as trustworthy as the numbers under it. Part 1 of a series on data foundations for enterprise AI: getting every channel's data in, and modeling it so it agrees with itself.Data Architecture
  3. 4 min readData models agents can trustAnalysts fill gaps in a data model with tribal knowledge; AI agents fill them with confident guesses. Four ideas, with real Salesforce examples, for models an agent can answer from correctly.AI NativeData Architecture
  4. 9 min readMigrating a chart of accounts without breaking a single reportWhen finance re-cuts the general ledger, every dbt model, metric and dashboard downstream is exposed. A playbook: crosswalks, dual-running, zero-tolerance reconciliation, and treating the account hierarchy as a data contract.MigrationBusiness Logic
  5. 8 min readCertifying data models with an agent in the loopA production pattern for deciding whether a dbt model is fit for finance reporting: deterministic gates first, sandboxed tests against production data, ticket traceability, and a coding agent used only for triage.AI Native
  6. 8 min readTurning regulatory PDFs into agent-ready packsModels are bad at 60-page government PDFs and good at well-shaped data. A recipe for converting one into the other, worked through on the public US product-safety eFiling documentation.AI Native
  7. 10 min readEvals for analytics agents: grading answers against the ledgerIf an AI answers finance questions, its accuracy should be a number you can report. How to build a golden question set, compute ground truth from certified models, score numeric and tabular answers, and catch regressions in CI.AI Native
  8. 2 min readFive long bets on AI, data and financial infrastructureWhere I'm putting time and capital over the next decade, the signal I watch for each bet, and what would prove me wrong.AI Native
  9. 9 min readThe semantic layer is the API for AIText-to-SQL demos well and fails quietly on finance data. Governed metrics, declared grain and semi-additive measures are what make an analytics agent trustworthy, and the semantic layer is how you hand them over.AI NativeData Architecture
  10. 8 min readBasis points on a moving balance: modeling asset-based feesCustody and asset-management fees are charged in basis points on balances that change every block. Average daily balance, tiered vs. banded pricing, minimums, valuation cutoffs, and reconciling the model to the invoice.Business Logic
  11. 9 min readOne customer, five systems: entity resolution and survivorship for finance dataCRM, contracts, ERP, billing and KYC each have their own idea of who the customer is. How I model a canonical party, match records without guessing, pick survivorship rules per attribute, and prove the result is right.Data ArchitectureBusiness Logic
  12. 8 min readAs-of vs. as-was: bitemporal modeling for financial reportingLate prices, backdated deposits and restatements mean "what was the balance on December 31?" has two right answers. How to model both, and why auditors care which one you show.Data ArchitectureBusiness Logic
  13. 9 min readLedger-first: a data architecture for a crypto custodianEvery finance number should be derivable from one immutable, double-entry ledger. How to model custody, balances and valuation so that dashboards are just views and reconciliation is a test.Data Architecture

In the works

  1. 01

    I left a federally chartered crypto bank to sell Chinese open-weight models to Southeast Asia

    A working draft about a change in direction, the questions behind model selection, and the business work that comes after a benchmark.

  2. 02

    Sovereign by layer: what actually stays in-country when an AI agent touches your data

    A practical way to trace storage, retrieval, inference, telemetry and support before making a claim about AI data residency.

  3. 03

    Your text-to-SQL agent is lying to your CFO

    The query can compile and the answer can still misstate revenue. A worked example and an acceptance contract for finance-facing agents.

Things I've built

  1. Data engineering · at work · 2026

    Agent-in-the-loop data certification

    A Claude Code plugin that certifies dbt models against team data standards before they reach dashboards.

    Runs the same static manifest checks as CI, executes model tests against production inside throwaway temp-table sessions, checks that every commit traces to a ticket, and has an AI step triage failures into "fix" or "escalate". A green run applies the certification label automatically.

    • dbt
    • BigQuery
    • Claude Code
    • Python

  2. AI × regulation · 2026

    CPSC eFiling agent pack

    Turned a stack of US product-safety PDFs into an agent-ready corpus, plus a test harness for the Product Registry API.

    Page-level markdown, ~80 extracted CSV tables, JSONL lookups and an API contract JSON with jq recipes, so an LLM can reason over the spec instead of guessing. The harness (Pydantic v2, httpx, FastAPI mock) generates synthetic fixtures, redacts every report, and refuses live calls unless explicitly allowed.

    • Python
    • Pydantic
    • FastAPI
    • pytest
    • LLM tooling

  3. Company site · bilingual · 2026

    WayAnalytics ↗

    A bilingual (English and Chinese) analytics and e-commerce compliance studio, built as a monorepo with shared tokens and validated content.

    Astro marketing site plus shared UI, design-token and schema packages. npm run check gates every build on content, route, SEO, accessibility, link, budget and visual-contract checks. Public copy carries claim and evidence metadata, so nothing unverified ships.

    • Astro
    • TypeScript
    • Design tokens
    • GitHub Pages

  4. Personal tool · PWA · 2025

    Lazy Journal ↗

    A local-first life tracker: one row per day, a dozen habit chips, five languages, no server.

    Started as a single HTML file with React from a CDN; grew into a Vite PWA with a today-first mobile view, a spreadsheet-dense desktop archive, lazy-loaded GeoNames search and SheetJS export. Your data never leaves the browser.

    • React
    • Vite
    • PWA
    • localStorage

  5. Infra for one · 2026

    Personal agent fleet

    Self-hosted AI agents and bots, with their own token, skill and log management.

    A working lab for running open agent frameworks day to day: gateways, scheduled routines, skill libraries and per-bot budgets. Mostly an excuse to learn where agents break before they break at work.

    • Python
    • Docker
    • MCP
    • LLM APIs