Nextralume

Inspiration

Operational workflows often live in scattered SOP docs, PDFs, screenshots, and chat threads. When something breaks, teams spend time arguing about "what happened" instead of fixing the process. I wanted a tool that turns messy inputs into a clear workflow map, highlights risk and bottlenecks, and produces improvements you can actually act on.

What Nextralume does

Nextralume converts operational knowledge into a risk-ranked workflow and a shareable report:

• Analyze a workflow using a chosen template (the demo includes a "shift handover" scenario, but the approach is general).
• Visualize steps as a timeline map with risk indicators.
• Explain "why" each risky step is risky in plain language.
• Produce a prioritized issue list and concrete recommendations tied to the steps.
• Export results to shareable formats (HTML/Markdown) and optionally send a webhook payload.
• Save analyses as sessions so teams can revisit decisions and changes over time.

How I built it

• Next.js 16 (App Router) + TypeScript for the web app.
• Zod schemas to validate structured AI outputs and keep results deterministic.
• Prisma + PostgreSQL for session persistence and auditability.
• Provider abstraction for LLM calls (ERNIE), with a manual smoke test to verify real API connectivity in staging/local environments.
• Strong quality gates: lint, typecheck, tests, and build in CI, plus automated deploy to a VPS.

Challenges

• Database migrations drift: an empty migration folder caused schema mismatch. I fixed it with a safe Postgres migration strategy (defaults, backfills, indexes) and added a regression test to prevent empty migrations.
• Next.js async params and async rendering boundaries: I adjusted route/page patterns and stabilized tests by rendering async components correctly and mocking fetch responses deterministically.
• Safety and robustness: upload limits, structured error handling, and guarded webhook sending to reduce operational risk.

What I learned and what's next

I learned that AI becomes much more usable when the output is constrained by a schema and grounded in a workflow UI. Next, I plan to integrate OCR for real document inputs, expand templates, and refine the workflow visualization for even faster review and collaboration.

Built With

• ernie
• github-actions
• motion
• motion-(motion-for-react)
• next.js
• novita
• postgresql
• prisma
• react
• react-testing-library
• shadcn/ui
• tailwind-css
• typescript
• vitest
• zod