Samuel McDonnell

I build the reliability layer under AI features — evals, harnesses, and the measurements that show whether they work.

AI Engineer at RWA, Ireland. Most of what I do is the unglamorous half of shipping AI: making a model's behaviour observable, bounded and cheap enough to run every day.

Every project on this site carries either a measured result or a written reason there isn't one. That second case is the more common one, and hiding it would make the first case worth less.

All projects

2026–present

Professional · RWA

RWA MCP Server

A production MCP server splitting typed, audited tool access into separate customer and authoring surfaces.

Measured internally at RWA — not mine to publish No measurement

  • Python
  • Model Context Protocol
  • FastMCP
9 components
  1. Tool registry Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE
  2. Customer / authoring split Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE
  3. Typed schema validation Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE
  4. Per-tool scoping Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE
  5. Fail-closed access gate Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE
  6. Metadata-first query gate Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE
  7. Guarded query execution Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE
  8. Demo sites as a proving ground Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE
  9. Call audit log Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE

2025–present

Professional · RWA

RWA Customer Portal

A Next.js portal giving customers self-serve dashboards over their own data, with a staff console behind it.

Measured internally at RWA — not mine to publish No measurement

  • Next.js
  • React
  • TypeScript
  • Tailwind CSS
  • Drizzle ORM
  • Auth.js
  • Tableau Embedding API
8 components
  1. Tenant-scoped dashboards Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE
  2. Tableau embed layer Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE
  3. Staff console Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE
  4. Session and revocation Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE
  5. Tenant and module catalog Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE
  6. Notifications Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE
  7. GDPR tooling Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE
  8. Rate limiting Evidence: MEASURED INTERNALLY — NOT PUBLISHABLE

2025–present

Personal

Glas Intelligence

Multi-agent simulation that turns uploaded documents into stakeholder agents and runs policy scenarios.

No accuracy number yet — and that is the honest answer No measurement

  • Python
  • Flask
  • Celery
  • Redis
  • Zep Cloud
  • OASIS (camel-ai)
  • Vue 3
  • Supabase
  • Docker
  • Nginx
8 components
  1. Document ingestion Evidence: CODE
  2. Persona synthesis Evidence: CODE
  3. Simulation runner Evidence: CODE
  4. Quantitative analysis Evidence: CODE
  5. Monte Carlo engine Evidence: CODE
  6. Calibration guardrails Evidence: CODE
  7. Report agent Evidence: CODE
  8. Forecast scoring Evidence: CODE

Writing

All writing