I build control planes for AI agents.
Policy gates, approval workflows, immutable audit, and eval harnesses — the deterministic layer that makes autonomous systems deployable.
Founder & CTO, TechTide AI · Columbus, OH
Most agent failures aren't model failures — they're governance failures. I design the enforcement layer that sits between an agent and anything that matters: deterministic policy engines with zero LLM calls in the decision path, human-in-the-loop approval gates, database-enforced immutable audit logs, and eval harnesses that treat policy regressions as build breaks. If a system can't show you a log, it isn't governed.
| Project | What it governs |
|---|---|
| ClawKeeper | Agent finance platform where a deterministic policy engine runs tenant, capability, and approval checks before any LLM call. Payments default to $0 autonomous spend, and the audit log is immutable at the database layer — UPDATE or DELETE raises an exception. |
| techtide-harness-kit | Trust contracts as JSON Schema: every skill declares its risk tier, tool permissions, deny-by-default network egress, approval gates, and compliance control mappings. Ships a prompt-injection threat model with adversarial fixtures, enforced by ~28 CI validation gates. |
| CipherClaw | Multi-agent debugger with zero LLM calls in the analysis path — causal graphs, weighted root-cause scoring, and failure prediction you can't prompt-inject. 206 zero-mock tests, zero runtime dependencies, no network egress. |
| swarm-357 | 357-agent orchestrator governed by per-agent budget caps, a pattern-matched bash security gate, layer-level daily budgets with automatic model downgrade, a hard per-run spend ceiling, and correlation-ID structured telemetry. |
| agentrewind · private | Time-travel debugging for agent traces: zero-trust import parsing (signature verification, size clamps, sanitization), five trace formats auto-detected, and a divergence/latency/token diff engine for side-by-side replay. |
| FintheFinder | Deep-research agent with a real human-in-the-loop approval workflow — the run suspends until a human approves — and a deterministic citation audit with zero AI in the verification path. |
| Molten · private | Governed agent runtime: kill switch with cost and error auto-triggers, circuit breakers, a four-tier execution sandbox (read-only → admin), hard daily and per-task budget caps, and a dead-man's switch. |
| TechTideAI2 | Company-scale agent platform on a three-plane architecture — control, data, and evidence. Status transitions run through an explicit state machine, execution has a hard timeout, and changes are graded by an adversarial eval harness with a 33-task golden suite. |
Logs over vibes.
Production over theater.
Deny by default.
The trustworthy part of an agent system is the part with no AI in it. Policy checks, budget ceilings, audit trails, and approval gates should be boring, deterministic functions — because you can't prompt-inject a function.
The same discipline, before it had a name: production marketplace infrastructure with idempotent payment webhooks, forward-only migrations, end-to-end trace propagation into an ops event log, and ten pre-written incident runbooks — on top of 13 years in US enterprise IT. Governance isn't a pivot; it's the pattern.
- TechTide AI — client engagements: Production Triage, Workflow Rescue, Fractional FDE
- techtide-harness-kit — trust contracts and CI gates for agent skills
- swarm-357 — governed multi-agent orchestration
Updated quarterly.
- Automation Vibes — podcast + newsletter with Shane Spencer
- Anthropic Partner Network — services partner
Production Triage available. Book in the Featured section on LinkedIn or at techtideai.io.
