Four things I can point at

Agent toolchain · maintained fork · TypeScript · public

codex-superpower

A maintained fork of a multi-model coding-agent toolchain. I ported roughly thirty fixes from upstream PRs and other forks — each with its own regression test, each on its own branch: account-wide rate-limit backoff, compaction handling during limits, stricter origin checks for the local bridge.

The installer builds from source and runs the full test suite before any update; if the new build fails to start, it rolls back to the previous one. Self-updating tooling that cannot brick itself is a product requirement, not a nicety.

github.com/trukhinyuri/codex-superpower

Boundary: it is a fork with a small community; I am not the upstream maintainer. Treat it as maintainership practice of an agent toolchain, not as adoption.

Agent infrastructure · Python + Go · private core, public shell

agent-tools & CloudRING

agent-tools (private)

The operating pattern of my agent fleet: a single read-only gateway through which AI agents reach external systems — mail, trackers, chat, docs. One audited path with per-tool boundaries, instead of credentials scattered across harnesses. In daily use across machines and agent harnesses since 2026; actively maintained.

CloudRING (public)

The public core of an open-core cloud platform structure: Go, Apache-2.0, 125 commits as sole contributor. PR #163 hardened a negative TLS test — merged with automated checks green. Around it sits a private product layer with reference deployment, gitops and governance documents.

github.com/opencloudtech/CloudRING

Boundary: CloudRING’s own README marks it evaluation-stage — not production- or pilot-ready — and no external adoption is claimed. agent-tools is private: the pattern is public, the internals are not.

Quantitative research · Go + CUDA · private

capital3

A private quantitative research project: numerical methods running on GPUs, built as a long-running personal system in Go and CUDA. It keeps my hands in systems programming and numerical computing — the same instincts as platform work, minus the committee.

Boundary: the repository is private and its results are unpublished — treat it as direction, not evidence.

Cloud product · NVIDIA V100 / vGPU · 2019–2020 · historical

The MTS GPU product

A public cloud GPU service line, from zero: configurations (dedicated V100 instances and vGPU profiles), tariffs, availability design, and client scenarios across training and inference. I owned the product decisions — what to sell, at what price, with what promise — working with the engineering teams that ran it.

Public record: my AI Journey 2020 talk on GPU clouds. What that experience taught me about infrastructure economics is the subject of the first post on this site.

Boundary: a historical product. Client-effect figures from that period are not cited; the talk’s forecasts were forecasts, not achieved metrics.