AI Software Engineer

Software engineer with 15+ years of experience, frontend at the centre since 2016. Today I ship mostly through agents: I write the spec, a governed loop of Claude Code, Codex and reviewer agents delivers it, and I rule on what the gates flag.

Background Production software across SaaS platforms, real-time systems, cross-platform applications and Web3, owning architecture end to end. React since 2016, Next.js since 2020. At Game7 (November 2023 – March 2026) I was Senior Frontend Architect: a multi-tenant platform with swappable design systems behind enforced type contracts, and multi-agent content generation with model routing matched to task complexity. Independent since 2026. Four years working with AI every day.

How I Work I write the spec; agents ship the feature. Every change passes deterministic gates on my machine and in CI, then two isolated reviewer agents judge the same frozen input. Every verdict is bound to the exact commit it judged, and when the loop can't converge it stops and asks me. I don't read every line agents write. I design what every line must pass, and I make the calls the gates can't.

Projects Vinaya — a governed delivery loop for coding agents. The loop above, built by me. Claude Code or Codex develops inside an OS sandbox; deterministic gates and two reviewers (code and security) decide; any push or rule change voids a verdict, so a stale approval can't reach merge. 589 merged pull requests since July 2026. Every governed act lands in a typed event log, streamed live. Now moving onto my own LangGraph engine as a typed, checkpointed state machine. → vinaya.attalabs.dev

Onchain Rewind — an app built entirely through Vinaya. Every ticket planned, coded by an AI developer, checked by two AI reviewers, then merged. Its public dev log shows the real numbers, drawn from GitHub and the Vinaya log: 32 of 33 tickets merged, a median of 34 minutes of agent time per ticket, 103 problems caught before merge. → zerion.attalabs.dev/dev-stats

Vāda — multi-agent deliberation engine. Agent teams written as YAML specifications and compiled into executable LangGraph flows, routing across Claude, Gemini, GPT and Grok with cost, quality and fallback logic. Hosted Model Context Protocol (MCP) server with bearer auth and envelope-encrypted keys. → vada.attalabs.dev

Herald — this page. Forensic CV-to-job match audits, multi-tenant with per-user theming from Sanity CMS.

Interfaces The part I care about most: making agent work readable. A live agent-execution graph animated from streamed events, streaming chat and prompt composers, Vinaya Studio, and Mission Control, a live view of every agent run I designed so that a dead run never looks paused and unknown never shows as zero. All of it sits on a design system of four interchangeable libraries held to one enforced contract.

All of it lives in Atta Labs, my lab, and Vinaya governs how every line of it gets built.