I build verifiable AI — engines, tools, and products where every claim is checked before it ships.
I don't guess. I don't bluff. Everything ships through the same gate: deterministic where I can make it, audited everywhere I can't, and never fabricated. Tests green or it's not done.
Truth is a consensus — not one model's assertion.
verify → remember → recompile → consensus
hybrid-agi-vre wraps language models so every inference step is checked before it's committed:
- deterministic verification — same input, same signature, every time
- gated self-improvement — patches its own code behind a tested circuit breaker + rollback
- distributed consensus — contested claims settled by a weighted vote of independent nodes, not one confident answer
AGPL-3.0. Runnable package + installable Hermes plugin (tools + /pillars-audit).
The tokenizer is a formal computational primitive grounded in 3,000-year-old text — used as math, not mysticism. No learned weights on the verification path. Reproducible forever.
A family of deterministic, no-LLM-in-the-critical-path tools, all E2E-tested and one-command installable:
| category | repo | what it does |
|---|---|---|
| ⚡ flagship | mcp-token-saver |
request-path token proxy + analyzer — cut 74% off a live DeepSeek call (live endpoint: mcp-token-saver-proxy.fly.dev) |
| security | mcp-skill-sec |
audit skills/prompts vs 8 malicious-skill supply-chain patterns |
| security | mcp-verify-claim |
evidence-gated FACT/INFERENCE/SPECULATION reporting |
| security | mcp-benchmark-hygiene |
detect pytest config-leakage corrupting agent-eval grading |
| security | mcp-secret-scrub |
secret scrubber for agent output |
| security | mcp-auth-audit |
auth-security config audit |
| CLI / eval | harness-audit |
audit an agent-eval workspace for the silent mis-scoring bug class (the $2–8K service lead-magnet) |
| CLI | cov-shield |
run pytest with coverage-gate leakage neutralized (fixer companion) |
| CLI | env-precedence-check |
detect CLI-default-silently-overrides-env-var bugs |
| CLI | ci-diff-audit |
audit what a pipeline changed vs declared intent |
| CLI | mcp-schema-lint |
validate MCP server.json + tool surface against the Official Registry schema |
| data | token-analytics |
deterministic agent-token-cost intelligence from real usage data |
Multiple are published on the Official MCP Registry and PyPI. All MIT (except the AGPL core).
Fresh data product: token-analytics — reproducible JSON reports on real agent token cost (our own stack: 962 req/day, 289K tok/req, 99.22% cache-hit, $29/mo reclaimable cache-miss tail).
I build production MCP servers and agent-verification layers at fixed price, milestone-billed ($8K–$50K). Full scope table + engagement path: agensi-builds.
Python · NumPy · Hermes Agent · MCP · deterministic audit · gematria · gated recompilation · swarm consensus
The method is open. Knowledge should be free — the verification should be rigorous.