Most of my ~200 repositories are private, so there is not much to see here. Hit me up if you're curious.

• CTO at Group Advisors — a US employee-benefits platform (wellness programs, affiliate and employer portals, payroll integrations, billing, e-signature, an employee mobile app).
• An Agentic OS — my own operating system for software delivery run by AI agents. A controller seat plans and promotes, a warden answers and routes, worker lanes claim tickets from an in-house board, and a deterministic pipeline merges, builds staging, verifies and promotes to production. Dozens of agents work in parallel on one Linux box, around the clock, with humans only answering real decisions.
• Also active across my other companies: NeoGen.AI, COVID Schedule, VentureVoIP, CreateOffshoreCompany, SineApps, VentureIP, C O International Holdings, Bio Earth Farms, Light Stream Farms and Singularity Software.

_{The fleet dashboard I built for the Agentic OS — click a thumbnail for the full screenshot. Top row: session control and the delivery pipeline · fleet load, account rate limits, machine vitals and open questions · blockers by person, backlog, completions and workers vs RAM. Bottom row: the usage page that lives on a separate small HDMI screen · the same overview in the Nebula and Tron themes. Ticket, PR, question and account details blurred.}

We stopped paying for hosted CI runners (GitHub-hosted minutes first, then Blacksmith) and replaced the whole path with our own:

• A merge-ready FIFO that agents hand finished PRs into, instead of a hosted merge queue.
• One combined batch every 30 minutes: rebase the queued PRs onto fresh main, run a single compile / lint / changes-only test / catalog / help gate, merge the integration branch.
• One staging build from a source tarball via the Heroku Build API, verified by served commit and health, then promoted as the exact slug to production. Production is a copy, never a second build.
• Automatic triage: a failed gate is attributed to the culprit PR, which is held with the exact error while the rest of the batch relaunches without it.
• Self-hosted runners on our own Linux box for the few scheduled jobs that still use GitHub Actions; the security scans, nightly suites and hourly full test runs are systemd timers.
• Everything writes a machine-readable verdict line; nothing is "green" without evidence.

I don't use AI as an autocomplete; I run it as a workforce.

• Background: I have built and fine-tuned my own models from scratch, and I have extensive experience deploying ensemble models in production — the approach that has since been popularised as mixture of experts.
• Models: Anthropic Claude (Fable / Opus 5 for control and judgment), Zhipu GLM 5.3 for volume work, OpenAI Codex models, xAI Grok, plus local models via Ollama.
• Tools: Claude Code, T3 Code, opencode and Codex CLI as agent runtimes; systemd timers for the deterministic parts; Telegram for the human loop; Playwright for browser verification; self-hosted GitHub Actions runners; Heroku for delivery.
• Principles: everything in production is verified by evidence, not by a green check; agents claim work with locks and audit trails; every decision is recorded with who said it and when; and anything that is done twice by hand becomes a script or a timer.

TypeScript · React · Vite · Tailwind · LoopBack 4 / Node.js · MySQL · Heroku · Expo / React Native · Firebase · Playwright · Jest · GitHub Actions · Linux / systemd · Python