© 2026 Medoya LLC
Powered by voilamia
Subscribe to receive new insights
You can't responsibly advise on AI without building it. These are the tools I've built to find out what holds up in production: agent guardrails, tools, benchmarks and MVPs. Most are open source, and each one draws on 25 years of software architecture, AI and security work.
The engine under the sites I ship and my own ongoing test bench: a proven headless CMS underneath, enforced layering above it, a permission floor no AI credential opens, and personal data encrypted field by field.
Jev 1.13 from TypeSafe against fast LLMs on MedHallu, next to Stanford's published MedHELM results: accuracy, latency and cost in one table, with run plans written before each run.
Reference architecture presented at ISC2 Security Congress 2025: Kong for access, OPA for risk-based policy, and a checker model that vets the generator, down to prompt injections hidden in medical images.
Every local engine on every machine you own, pooled behind one OpenAI-compatible address: your agents get one key with a budget, your security lead gets one place where access and spend are decided, and patient data stays on your network.
Local proxy for autonomous agents: proxy tokens instead of real keys, hard budget limits, and an instant kill switch for rogue connections.
CLI that generates OpenTelemetry logs, metrics, and traces with mock data and sensitive-data flagging, so you can test PHI scrubbing before real data flows.
Natural language to SQL over PostgreSQL: ask the data a question, review the generated query, see the chart.
Azure chatbot over GTFS static and real-time feeds, grounded in the official schedule so it answers only from authoritative data.