Lead Software Engineer · Deterministic infrastructure for AI agents · Ontology graphs · MCP tooling · Spec-driven SDLC · Evals
I build the deterministic layer that lets AI agents do real engineering work inside a real company. Every AI coding tool fails the same way at scale: not at writing code, but at knowing the organization — which service consumes which event, what breaks three repos away when a column is renamed, which rules are non-negotiable. I stopped asking models to guess at that and compiled it instead.
The thesis: production-grade AI engineering needs pre-computed structural understanding, deterministic retrieval, and machine-verifiable workflows — so agents operate on facts, not vibes. Models are for judgment. Everything else should be looked up.
- An ontology graph of the whole company — code, database schema, infrastructure, frontend, mobile and product decisions compiled into one graph. Typed query operations answered deterministically, with no model call. Every answer traces to source, and drift verification re-checks it against live code before anyone acts on it.
- A governed delivery pipeline — a Claude Code plugin that drives spec → plan → blast radius → parallel agents in isolated worktrees → adversarial review → PR, with an engineering constitution loaded on every run.
- Enforcement that proves itself — lifecycle hooks, CI gates that each ship with a negative self-test, structured error envelopes so agents self-correct, and structure-only telemetry with per-field redaction.
- Background agents that cannot merge — deterministic monitors and spec-gated builders emit draft PRs only; the merge capability is physically removed.
▶ The platform in three minutes · 📄 Full story, principles and writing
| Project | What it does | Install |
|---|---|---|
| skills | 10 skills + 18 agents for Claude Code, Codex, Cursor, OpenCode and 9 more harnesses, symlinked from one clone. Includes visual-verify: the agent looks at what it built before claiming it works. |
npx @vimoxshah/skills |
| tokenflow | See where your AI tokens actually go. Local-first analytics across Claude Code, Codex, Cursor, OpenCode, Cline. Zero dependencies. Live demo. | brew install --cask tokenflow |
| claude-session-replay | Turn an agent session transcript into one self-contained HTML player you can scrub, step through and share. | npx claude-session-replay session.jsonl |
| claude-router | Route each task to the cheapest Claude tier that can do it well, via model-pinned subagents. Only write lanes touch code; every diff is verified by a different model. | npx @vimoxshah/skills --bundles routing |
| claude-codex-orchestrator | Claude plans and verifies, Codex executes. Bounded work packets, three executor profiles, an escalation path. | npx @vimoxshah/skills --bundles routing |
- How I made a knowledge graph that checks itself
- Designing an MCP surface under the attention ceiling
- The agent that cannot merge
Open to conversations about agent infrastructure, evals and deterministic retrieval — email or LinkedIn.