Hey HN, I'm the founder of BrewHub. We run physical coffee and parcel lockers, and we deployed an autonomous AI concierge using Next.js and Python (Cloud Run). We quickly realized that validating LLM tool arguments in a polyglot setup was a massive security gap, especially since we were fighting AWS Lambda's 4KB environment variable ceiling for our secrets. I wrote up a formal engineering case study on how we built 'Parity Contracts' to enforce CI-blocked safety gates across runtimes before an agent can write to Square or Supabase, plus our Doppler disk-fallback hack. The red-team corpus is open-sourced on GitHub. Happy to answer any questions about the architecture!
Hey HN, I'm the founder of BrewHub. We run physical coffee and parcel lockers, and we deployed an autonomous AI concierge using Next.js and Python (Cloud Run). We quickly realized that validating LLM tool arguments in a polyglot setup was a massive security gap, especially since we were fighting AWS Lambda's 4KB environment variable ceiling for our secrets. I wrote up a formal engineering case study on how we built 'Parity Contracts' to enforce CI-blocked safety gates across runtimes before an agent can write to Square or Supabase, plus our Doppler disk-fallback hack. The red-team corpus is open-sourced on GitHub. Happy to answer any questions about the architecture!