I'm Travis, a staff engineer with ~20 years building software. Eleven months ago I started building Flywheel — an AI workflow automation platform. Solo. Evenings and weekends.
Today the codebase is 325,000 lines of production code. 702 commits. Three services (Go backend, Next.js frontend, Python AI agent), 14 integrations, full CI/CD, Terraform infrastructure. Enterprise-grade auth, billing, monitoring — the works.
I didn't mass-generate code. I didn't skip tests. I wrote a real system that I'm shipping to real users. Here's how.
The Stats
Let's get specific. These are real numbers, not estimates.
Codebase:
- 170K lines of Go (backend API, 14 provider integrations, distributed pipeline execution)
- 145K lines of TypeScript (visual workflow editor, React Flow canvas, 60+ component directories)
- 7.5K lines of Python (AI agent — FastAPI, Anthropic SDK, MCP tools)
- 2,500 lines of Terraform (GCP infrastructure as code)
- 847 test files across the stack
Git churn:
- 1.37 million lines inserted
- 716K lines deleted
- Only 325K survived
- For every line of code in the app, I wrote more than two
Claude Code usage (all time):
- 2,900 sessions across 144 active days (longest streak: 42 days straight)
- 94.7 million tokens generated — plus billions more served from prompt cache
- ~$1,200 in Claude Max subscription — usage that would've metered at $2,500+ on API pricing
What I Actually Built
Flywheel automates recurring tasks. Not one-off AI magic — persistent workers that run on schedule, via webhook, or triggered by events.
The core insight: AI is great at reading messy data (emails, PDFs, Slack). Deterministic logic is great at routing, filtering, delivering. Combine them and you get workflows that handle jobs reliably.
What the platform does today:
- Visual drag-and-drop workflow builder on a React Flow canvas
- 10 data providers: Gmail, BigQuery, PostgreSQL, S3, DynamoDB, Firestore, GCS, Google Drive, Pub/Sub, Domo
- AI enrichment nodes: Claude and ChatGPT run mid-pipeline, reading data and producing structured output
- 8 processing transforms: filter, route, join, collect, flatten, split, apply template
- Human-in-the-loop: workflows pause for human review before continuing
- Distributed execution with chunked streaming, per-node error policies, failure retry
- Multi-org RBAC with entity-level permissions
- Stripe billing, monitoring dashboard, template marketplace
The workflow canvas isn't just where you build. It's where you run your operation. That's the 1.0 vision — you open the canvas and see live stats, reports, and action queues from your running workflows, not just the pipeline graph.
Why the Fundamentals Post Matters More Now
Six months ago I wrote about how . Want to see what a real workflow looks like? The that puts petty team grievances on trial in a dedicated courtroom channel, complete with an AI judge — because dogfooding should be fun too.)
Has the ceiling for what one person can build actually changed? Or am I just in the honeymoon phase? Curious what others are seeing.
SOCIAL SHARE CARD GENERATOR