Live SaaS · founder, engineer, designer
LaserBurnAI
A production SaaS that turns images into laser-ready vector files, with a metered credit system, five free browser tools, and 1,300+ automated tests.
- Status
- Live, paying customers
- Period
- 2026 → present
- Role
- Solo founder: product, engineering, design, growth
01 — Problem
The problem
Laser engravers and cutters need clean vector paths. AI image generators produce rasters. Getting from one to the other meant a manual chain of Illustrator, Inkscape, and guesswork, and most hobbyists gave up before the first cut.
The business constraint was just as hard as the technical one: every generation costs real API money, so the product could not give anything metered away for free.
02 — Build
What I built
- An image-to-vector pipeline: prompt-driven or uploaded images are traced server-side into layered SVG with separate cut and engrave layers, plus DXF R12 export for older laser software.
- A credit system backed by a ledger table. Every deduction, refund, and grant writes a row in the same transaction as the balance change, so a failed generation refunds atomically instead of silently eating a credit.
- Idempotent billing webhooks: a partial unique index on the order id means a replayed LemonSqueezy event returns 200 and grants nothing twice.
- Five free, client-side tools that cost nothing to run and bring in organic traffic: a layered slicer for 3D models, sheet nesting, a parametric box generator, single-line art, and an image-to-heightmap converter for CNC relief carving.
- GA4 purchase attribution end to end, a support intake with a plan-aware auto-triage gate, per-route rate limiting, and a test suite that grew with every incident.
03 — Decisions
The decisions that mattered
Anyone can list a stack. These are the calls that changed the outcome.
Engrave-first, not cut-everything
Early versions tried to cut every interior shape. Real users wanted a clean engraving with an outer silhouette cut and interior cutouts opt-in. The tool now defaults to the thing people actually burn.
Zero free credits, by design
Production data showed most sign-ups never generated anything. A free tier would have spent API money on people who were never going to buy. Free tools run entirely in the browser instead, which is where the organic traffic comes from.
A ledger, not a balance
The first refund bug wrote the already-decremented balance back, a no-op that charged people for failures. Moving to append-only transactions with RPC-only writes made the money trail auditable and the bug class impossible.
Platform timeouts are not exceptions
A serverless kill at 60 seconds bypasses every try/catch. Long generations were restructured so a timeout cannot leave a charge without a refund.
04 — Result
Result
automated tests, mutation-checked on new routes
live tools: 2 metered, 5 free and client-side
organic paying conversions at launch
05 — Next
What I'd do next
- Move the trace step into a browser Worker so the paid tool gets the same zero-marginal-cost profile as the free ones.
- Finish the physical test cuts for the parametric tools; the geometry is verified in software, not yet on a bed.