Ron Bronson
Work · About · Coaching · Contact
← Back to all work
delivery

Occupant — AI Compute Benchmarks

Public data infrastructure for AI model pricing, compute deflation tracking, and market intelligence.

Occupant — AI Compute Benchmarks

Models & Artifacts

Occupant produces several public benchmark models:

1. Compute CPI (Consumer Price Index)

A price deflation index tracking how fast AI compute costs are falling across capability tiers. Answers: “Is compute getting cheaper? How fast?“

2. AI Economic Activity Index (AEAI)

Market activity proxy measuring economic growth in AI usage. Tracks volume, diversity, and market structure changes.

3. Market Intelligence Dashboard

Sabermetrics-style analytics showing:

  • Model rankings and performance tiers
  • Premiumization vs. commoditization signals
  • Provider market share and pricing strategies
  • Price volatility and discrepancy alerts

4. Government Procurement Benchmarks

Fair market price ranges and vendor evaluation questions specifically for public-sector buyers. Includes procurement worksheet templates.

5. Cost Calculator

Interactive estimator for budgeting AI spend based on usage patterns and provider pricing.

The Problem

AI model pricing is volatile and fragmented across providers, resellers, and model routers. Buyers can’t easily answer:

  • What’s the fair market price per token for a given capability?
  • Is our vendor giving us a competitive rate?
  • Should we budget for continued deflation or stabilization?
  • What procurement questions should we ask?

The Solution: Public Data Infrastructure

A static website that aggregates pricing from multiple sources (OpenRouter, LiteLLM, Chatbot Arena, etc.), computes indices daily via GitHub Actions, and publishes JSON datasets + interactive dashboards.

Key design decisions:

Static Site + Daily Regeneration

No database. No server. Just:

  • Daily cron job fetches pricing data
  • Python scripts compute CPI, AEAI, market intel
  • Regenerates JSON files committed to repo
  • GitHub Pages deploys updated static site

Client-Side Rendering

Calculator and dashboards use vanilla JavaScript + Fetch API. No framework bloat. Progressive Web App (PWA) with service worker for offline access.

Zero Tracking

  • No analytics
  • No cookies (except localStorage for theme)
  • No user data collection
  • Short-retention server logs only

Data Sources & Credibility

Aggregates from multiple public sources with cross-checking:

  • OpenRouter Rankings (Weekly)
  • Chatbot Arena (Weekly)
  • LiteLLM pricing feeds (Daily)
  • pricepertoken.com (Daily)

Price discrepancies exceeding 10% trigger manual review before publication.

Full transparency disclosure published at /tardigrade.html covering data sources, AI usage, and privacy posture.

Technical Architecture

Backend (Python + GitHub Actions):

  • src/data_collector.py - Fetches pricing from all sources
  • src/calculate_cpi.py - Computes CPI baskets and deflation rates
  • src/calculate_aeai.py - Computes activity index
  • src/market_intel.py - Generates rankings and signals
  • src/scrape_rankings.py - Pulls OpenRouter/Arena data

Frontend (Static HTML + Vanilla JS):

  • index.html - Landing page with live index values
  • cpi-data.html - CPI dashboard + historical data
  • aeai.html - AEAI dashboard
  • sabermetrics.html - Market intelligence
  • calculator.html - Interactive cost estimator
  • gov.html + gov-worksheet.html - Procurement tools
  • glossary.html - Metric definitions
  • tardigrade.html - Transparency disclosure

Data Artifacts: All datasets published as JSON under /data/ for programmatic access.

User Journeys

Procurement / Finance User:

  1. Checks CPI/AEAI headline numbers on index
  2. Opens gov benchmark page for “fair deal” ranges
  3. Uses calculator to estimate spend
  4. Downloads worksheet for vendor evaluation

Research / Market Watcher:

  1. Reviews CPI time series for deflation trends
  2. Opens market intel for structure analysis
  3. References glossary for metric definitions
  4. Accesses JSON API for analysis

Developer / Analyst:

  1. Fetches JSON datasets directly
  2. Integrates into internal dashboards
  3. Uses data for budgeting models

Why It’s Built This Way

  1. Static = resilient, cheap, fast

    • No server to crash or scale
    • GitHub Pages hosting (free)
    • Cacheable, works offline (PWA)

  2. Daily regeneration = fresh data without live queries

    • Compute happens in CI, not at request time
    • Users get instant page loads
    • No API rate limits to worry about

  3. Public data infrastructure

    • No login required
    • No tracking or paywalls
    • Transparent methodology
    • Downloadable datasets

What This Demonstrates

  • Public data infrastructure design: Building credible benchmarks that serve procurement, research, and strategic planning
  • Static architecture patterns: Proving you don’t need a database/server for daily-updated data products
  • Transparency posture: Publishing methodology, sources, and AI usage disclosure
  • Progressive enhancement: PWA features for offline access without JavaScript requirements
  • Government-focused UX: Procurement worksheet and “fair deal” benchmarks tailored to public sector needs