Swiftly
Resume automation: parse once, autofill anywhere, compile in a sandbox
- 100+
- resumes processed
- 3
- form frameworks autofilled
- 1GB/1CPU/60s
- LaTeX sandbox budget
The problem
Job applications ask for the same twenty facts in twenty different shapes. The data already exists in a resume PDF; the work is getting it into a structured form once, then back out into whatever form a careers site happens to render.
Approach
Parsing that fails loudly
Gemini 2.5 Flash turns PDF, DOCX and LaTeX resumes into JSON validated against a schema before it touches the database. Normalised PostgreSQL via Prisma means a profile is relational data, not a blob — so a single employer rename updates everywhere it appears.
Two-stage autofill
Stage one extracts DOM field metadata from the target form; stage two maps that metadata to profile values. Injection is event-safe for React, Vue and Svelte, because setting .value on a controlled input is a no-op unless you dispatch what the framework is listening for.
Auth built, not bought
No third-party auth provider: JWT with refresh-token rotation, Google OAuth2, and WebAuthn passkeys. Rotation means a stolen refresh token is single-use and detectable on reuse.
Treat LaTeX as hostile input
User-supplied LaTeX is arbitrary code, so compilation runs in an internal Docker service capped at 1 GB / 1 CPU with a 60 s timeout and per-job temp directories. CI/CD gates on type-check + lint, pushes SHA-tagged images to Docker Hub and deploys the pinned SHA over SSH to EC2 as health-checked, non-root containers.
Outcome
- 100+ resumes processed from testing into production use.
- Autofill works across React, Vue and Svelte career portals without per-site rules.
- Deploys are pinned by image SHA, so a rollback is a redeploy of a known-good tag.
What I'd do differently
Hand-rolling auth was the right call for learning and the wrong call for velocity. The passkey flow alone cost more than the parsing pipeline — worth it once, not worth repeating on the next product.