Scratchpad / Clearskies
Senior Software Engineer — present · RemoteFull-stack engineer, Clearskies agent & workflow platform · built its frontend from scratch
Senior Staff at ZenGRC; Scratchpad has no engineering levels above Senior. I joined to build the Clearskies platform: I write specs, architect major features, and I'm the #1 contributor on both the frontend and the backend.
#1 contributor to both the Clearskies frontend (Vue 3/TypeScript/Electron) and backend (Go/Temporal/Postgres/MCP): 1,705 merged PRs and ~990 reviews across six repos in two years.
Scratchpad builds AI tooling for revenue teams; Clearskies (clearskies.cc) is its agent and workflow platform — 'the context layer for revenue AI' — reachable from Claude, ChatGPT and Gemini over MCP.
- Clearskies Workflows. Architected and built most of Clearskies Workflows, the product's core automation engine, end to end across a Vue Flow canvas and a Go/Temporal backend — growing it in ~8 months from CRUD to a graph-driven execution engine with 10+ node types, nested/parallel loops, scheduled and CRM-change triggers, dry-run test runs and versioned drafts. Replaced the hardcoded trigger→agent→Slack sequence with topological execution over the workflow DAG, so any node can run anywhere in the graph.
- AI Workflow Builder. Built the natural-language AI Workflow Builder — an LLM agent that generates and edits workflow graphs from chat and Slack — and made it reliable with publish-time validation and self-repair loops, an LLM semantic auditor, and a staged migration to native tool calling routed through the MCP server; then led its retirement (−28.5k lines) in favor of a headless builder running inside customers' own Claude and ChatGPT.
- Agent evals. Stood up LLM evaluation for the builder agent — from a Braintrust harness to an in-house framework with structural scorers, an LLM user-simulator for multi-turn disambiguation and a run-comparison UI — and gated prompt and tool changes on it.
- Human Review. Wrote the spec for and led Human Review: agent actions that touch Salesforce, Slack or email pause durably on Temporal until a named reviewer approves them, with reviewer notifications and production monitoring — ~70 PRs across frontend and backend in five weeks.
- Agent technical evaluations. Led technical evaluations of emerging agent approaches, including sandboxed coding agents (Codex, Claude Code), with a focus on credential isolation and least-privilege tokens.
7 more
- Clearskies frontend foundations. Bootstrapped the Clearskies frontend from scratch — chose the stack (Vue 3, Pinia Colada, shadcn/Reka on Tailwind 4, Vitest, pnpm), set team conventions, and wrote the agent tooling and CLAUDE.md guidance the team codes with.
- Multi-provider LLM layer. Turned an OpenAI-only LLM layer into a multi-provider platform (Anthropic, Gemini, bring-your-own-key) with an automated model-catalog sync and spend reconciliation against provider invoices, which caught and fixed a gap in recorded LLM cost.
- MCP security. Hardened the MCP server customers' AI assistants use to reach CRM context: server-enforced token scopes so workflow AI tokens can't call write tools, MCP tokens blocked as app logins, and an error contract designed for LLM clients.
- Deterministic meeting signals. Replaced costly, nondeterministic LLM meeting filters with deterministic attendance signals normalized across seven meeting providers (Gong, Teams, Fathom, Granola, Fireflies, Recall, Scratchpad), gating the production migration on an offline backtest and live shadow comparison.
- Frame AI drafts. On Clearskies' new product Frame (frame.page), became #2 contributor within a month (92 PRs): wrote the commenting and notification specs, built an AI drafting system (streaming rewrites, multi-version compare, per-change accept/reject, metering) and personally ran its prompt-injection security review.
- Failure alerting. Gave silent workflow failures an owner: classified, deduplicated and rate-limited failure alerts with fix links to the failing node, Datadog metrics, and runbooks/monitors for review, messaging, email and the model catalog.
- On Scratchpad's core app (2024–25), led frontend for Ask AI — LLM chat over deal, call and email context with streaming and transcript-timestamp citations — and drove a dead-feature cleanup that deleted ~116k lines.