Operator. GM & Product, in fintech and regulated industries.

This GitHub is where I keep small, opinionated tools I build to think clearly about applied AI, frontier science, regulated systems, and decisions worth pressure-testing. Code is intentionally simple; the value is in the framing.

Applied AI in regulated systems

• clinvar-classify — Research probe: LLM-assisted clinical variant classification with retrieval bundles. 60 ClinVar variants, no training.
• kyc-validator — Local-only KYC document validator. LLM extraction + sanctions screening, all on-device. PII never leaves the machine.
• stablecoin-policy — Selective-disclosure policy engine for stablecoin payments. Per-role views, predicate attestations, declarative policy language. TypeScript.
• machine-payments — Agent-to-agent payment network with strict budgets, escrow, and signed receipts across card and crypto rails.
• agent-debate — Multi-agent advocate/skeptic/judge debate for any consequential decision: hiring, product, policy, M&A.

Evals & frontier AI

• fin-evals — Domain eval harness for LLMs on financial reasoning. Four narrow tasks scored deterministically.
• bio-bench — Narrow eval harness for LLMs on biological reasoning. Five task families, 80 examples.
• prompt-eval-arena — Rigorous A/B harness for prompt iteration. Both deterministic and LLM-as-judge scoring, paired Wilcoxon, effect sizes.
• arxiv-radar — Weekly frontier digest for arXiv. Surfaces emerging vs receding research themes.

LinkedIn: linkedin.com/in/lucacosentino or cosentino.me