Self-learning AI systems, live in production

01Why I built these

Most “AI experience” is a wrapper around one model. I wanted production systems that exercise the hard parts — self-correction, evaluation, and multi-provider orchestration — so the patterns are battle-tested before a team needs them. So I built three, and I run them.

02Diamond AI — self-correcting ML

An MLB prediction engine with an auto-learning pipeline. Predictions auto-grade hourly; the system journals its own accuracy and retunes its models overnight from yesterday’s misses. There’s a public OpenAPI spec plus Python and Node SDKs.

• Hourly auto-grading and nightly self-tuning
• Accuracy journaling — the model that ships today isn’t the one that runs next week
• Public OpenAPI + worked-example SDK repo

03Vantage AI — adaptive ensemble

Self-tuning equity-market analysis built on a three-model ensemble — momentum, mean-reversion, and sentiment — with weights that shift as accuracy data accumulates. A nightly grading-and-insight pipeline and a shadow-compare evaluation harness keep it honest, with Claude commentary on every call.

• Three-model ensemble with adaptive, accuracy-driven weights
• Shadow-compare harness to evaluate changes before they ship
• A forecasting architecture that ports to any domain, not just markets

04AI vs AI — multi-provider orchestration

A live debate platform: two LLMs argue an audience-supplied topic in real time, streamed token-by-token over SSE. Same prompt, different models — OpenAI, Anthropic, and Gemini — so you can watch the reasoning diverge. It’s a working demo of provider-agnostic orchestration.

• Multi-provider orchestration across three model vendors
• Server-side token streaming via SSE
• Audience prompts feed the debate live

05The point

These aren’t demos that die in a notebook — they’re deployed, monitored, and public. They prove I can architect AI that improves itself, evaluates itself, and isn’t locked to a single vendor. That’s the applied-AI edge I bring to a team.

6What it adds up to