I design & ship production software end to end, from enterprise platforms to AI systems that run in the real world.

A real-time control plane for an autonomous trading-bot fleet.

Koleva Orakl (formerly rindbase-trade) runs a fleet of prediction-market bots through split signal and execution workers on Postgres and Redis. A Next.js dashboard streams state over SSE, and an MCP server lets AI agents inspect and change bots only through readiness gates and kill-switches. Wallet keys are encrypted per wallet, and arming a bot for live trading fails closed.

• Fail-closed live arming: readiness attestation before any bot goes live
• Scoped agent writes: audited MCP tools, per-wallet kill-switch, bulk actions
• Rebuilt in July 2026: 9 → 6 core containers, execution engine rewritten

The AI proposes and a human approves, through one code path.

An autonomous pipeline that can't act without a human sign-off.

Multi-tenant AI agents, with a human in the loop on every send.

Relay lets an agency run AI agents for its clients. Customer agents answer on a web widget, WhatsApp, Instagram and Shopify, and hand the conversation to a person when they cannot help. Operator agents research leads and draft outreach, but never send it: every outbound action waits for a human approval. Clients get a portal with a live feed and PDF reports.

• Tenant isolation: two agent runtimes, operator-only tools never reach the customer path
• Untrusted-input hardening: send tools are opt-in, internal addresses blocked, webhooks signed
• Built for concurrency: per-chat serial queues, Postgres-backed rate limiting shared by every service