Idealistic agents, real-world problems

Idealistic agents, real-world problems

I love building AI agents that are idealistic — agents aimed at problems that

actually matter, not benchmarks that impress reviewers for a weekend.

• Mining shortfall forecasting for real operations (MOIL-Manganese-AI)
• Research that verifies its own citations before it speaks (citecheck)
• Retrieval that can prove where every sentence came from (production-rag, colpali-rag)

The bar I hold myself to

1. It must run. A demo that only works on my laptop is a story, not a system.

2. It must prove itself. Evals, guardrails, and citations are part of the product.

3. It must be frugal. Real models on a 6 GB GPU, no expensive API key required to reproduce.

> Optimism is cheap. Shipping an agent that survives contact with a real user is not.

That gap — between a clever idea and a dependable agent — is the only place worth working in.