How I build

I use AI to interrogate concepts, news releases and product releases, to find trends and design opportunities, and to draft code, documentation and scripts for the pipelines I build. I use Claude, Meta AI, and Hermes (an agent in my own WSL, running an NVIDIA Nemotron model through OpenRouter), plus the occasional question to ChatGPT.

I make the design decisions, run everything on devnet, and check each claim against the chain myself. I decide what counts as proof: nothing goes in a deck or a demo without a transaction signature behind it. I run every command myself and paste the real output, including the failures, and several of my best decisions have come from the failures.

The boundary I hold

• AI is for velocity: drafting programs, scripts and documentation.
• Signing keys stay in local keypair files that are never committed to a repository, and no seed phrase or signing keypair is ever pasted into an AI chat.
• Before I publish a repository, I scan its git history for key material (the script).
• Rule adopted 18 October 2026: I mask API keys and tokens in command output before I share it.
• Where an AI draft and the chain disagree, the chain decides.
• Every on-chain result I publish points to a transaction signature. Limits are stated as limits, not as results.

AI drafts a lot. I run every transaction myself.

See it applied: Contributor stats · How Referee was built · Security checks. Back to ishvir.xyz