I work in the open at Zion-sewers — public infrastructure for trust in agents. Every repo answers one question: how do I know I am right?
- mcp-guard — adversarial fuzzer for MCP servers. 65 findings against Anthropic's reference servers, including silent shell-injection hits.
- sqltok — token-budgeted schema selection for Text-to-SQL. 97% gold-table recall at a 2,000-token budget on BIRD.
- agenteval-bench — pytest for agent outputs. Deterministic matchers, regression tracking, CI gates.
- protobuf-fuzz-guard — protobuf schema CVE scanning plus fuzz-harness generation in four languages.
The live Zion board. Every card, every office, zero hidden queues. This is where the work above actually happens.
More writing at daily-deep-learning.




