Software Engineer · AI Agent Reliability · Developer Infrastructure · Automation
I build systems that verify what software and AI agents actually did, instead of treating a successful response or command exit as proof. My work focuses on explicit trust boundaries, deterministic testing, safe tool execution, reliable recovery, and observable production behavior.
Founder of RecepioHQ.
Live, no-signup browser demo: Try Outbound OS's AI send-policy engine — change synthetic safety conditions and see exactly which checks return PASS, HOLD, or BLOCK. Runs in the browser with no actual messages sent.
Official MCP Registry: Outbound OS — active, v0.2.1, packaged as a public local-use OCI container at ghcr.io/recepiohq/outbound-os-mcp:0.2.1 (verified in the read-only GitHub Actions check). This is not a hosted public MCP endpoint or evidence of third-party adoption.
| Project | What it solves | Availability |
|---|---|---|
| Outbound OS | Evidence-gated outbound workflows: verify buyer fit, message claims, authorization, and delivery state before sending | Open source — MIT |
| Deep Seek Fix | Checks coding-agent outcomes against task contracts, execution evidence, and deterministic verification gates | Source-available; non-commercial license |
| Unified MCP Control Plane | MCP tool discovery with OAuth 2.1, policy-based execution, server-side authorization and redaction | Source-available; non-commercial license |
| Windows Maintenance Orchestrator | Policy-driven update plans, bounded recovery, reboot-safe state, and verification of real machine outcomes | Source-available; non-commercial license |
Try Outbound OS v0.2.1 in one command, without installing globally:
npm exec --yes --package=https://github.com/RecepioHQ/outbound-os/releases/download/v0.2.1/outbound-os-0.2.1.tgz -- outbound-os demo --compareRequires Node.js 22+. The command was tested on Windows: it demonstrates both a permitted simulated action and a blocked, unverified mailbox with zero real emails sent. It downloads the package from GitHub, not the npm registry. No Node.js? Try the interactive browser demo.
Explore agent verification without live model keys: Deep Seek Fix first-run guide.
- Social Radar Platform — React/Express/PostgreSQL and Capacitor Android architecture, tenant isolation, AI rule evaluation, and security-focused device ingestion. Alpha; physical-device launch gates remain open.
- MT5 AI Trading Runtime — Demo-only Python/FastAPI execution architecture with broker adapters, deterministic risk controls, persistence, and failure reconciliation. Real-money execution is hard-blocked.
Python · TypeScript · FastAPI · React · PostgreSQL · SQLite · MCP · OAuth · CI · AI Agents · Test Automation
- Observe the failure and gather reproducible evidence.
- Constrain unsafe or uncertain side effects behind explicit policy.
- Fix the smallest verified root cause.
- Test deterministic acceptance and failure paths.
- Report verified, failed, blocked, and untested outcomes separately.
I don't claim adoption, benchmark superiority, or a passing CI result without evidence.
- n8n debugging checklist
- n8n lead-intake dry-run workflow — synthetic, no live sends
- Supabase Auth redirect URL checker
- RAG evidence and restricted tool-call guards
I work on application debugging, integrations, backend systems, agent infrastructure, and reliable automation. Open a well-scoped GitHub issue for Outbound OS for project-related bugs or contributions, and consult each repository's licensing and security policy first.
For commercial engagements: