# acwx.net

## Author
Aaron Williams — analytics, data visualization, and product engineering leader based in Los Angeles. Background in journalism (The Washington Post, The New York Times, SF Chronicle) turned data viz engineering turned analytics product development. Currently at Netflix Ads.

## Site
Personal website and blog at https://acwx.net

## Blog
URL: https://acwx.net/blog/
RSS: https://acwx.net/blog/feed.xml

### Posts
- "I Spent Over $200 Teaching a Model What 'Clean' Means" (2026-04-08) — https://acwx.net/blog/i-spent-over-200-teaching-a-model-what-clean-means
  I built a side project over a few weekends around LA restaurant health inspections and tested three approaches to answering data questions with LLMs. The cost gap was wild — but the real lesson was about where business logic should live.
- "More of What, Exactly?" (2026-03-23) — https://acwx.net/blog/more-of-what-exactly
  Every exec says AI lets us do more with fewer people. But more of what? On the blind spot in AI-driven headcount cuts and what engineering leaders should build instead.

## Pages
- Home: https://acwx.net/
- About: https://acwx.net/about
- Cocina: https://acwx.net/apps/cocina — Toy agentic analytics harness for LA restaurant health inspections

## Cocina API (for agents)
Base URL: https://acwx.net/apps/cocina/api

### POST /apps/cocina/api/query
Full harnessed pipeline: classify intent → plan → resolve from KV → validate → narrate.
Request body: { "q": "your question", "model": "gpt-4.1-mini" | "gpt-4.1" }
Response: { query, metric, metricName, narrative, data, meta, validation, model, latencyMs, costCents, cached, stages }

### POST /apps/cocina/api/raw
Raw LLM response with no data harness (for comparison).
Request body: { "q": "your question", "model": "gpt-4.1-mini" | "gpt-4.1" }
Response: { query, answer, model, latencyMs, costCents, cached }

### GET /apps/cocina/api/health
System status: seed state, record count, date range, daily budget.
Response: { status, seeded, lastSeeded, recordCount, dateRange, budget }

### Rate Limits
- 20 queries per IP per hour on POST endpoints
- Daily spend cap: $5 (500 cents)

## Topics
AI adoption, analytics engineering, engineering leadership, judgment architecture, decision infrastructure, metric governance, cognitive offloading vs cognitive surrender, data products

## Newsletter
Subscribe: https://buttondown.com/aboutaaron
RSS: https://acwx.net/blog/feed.xml

## Links
LinkedIn: https://www.linkedin.com/in/aboutaaron
GitHub: https://github.com/aboutaaron
Email: hi@acwx.net