Problem
Operators and managers regularly receive raw spreadsheet exports (sales, ops, inventory, performance) that take hours to manually parse into something a leadership team can act on.
Constraints
- Output briefs must remain accurate and avoid LLM hallucination on numerical data
- Upload flow must handle messy real-world spreadsheets (mixed schemas, inconsistent headers, partial data)
- Hosting and inference cost has to stay viable at low usage volume during early access
Delivery
- Built the upload + parse pipeline on Next.js with Prisma over SQLite for fast solo iteration.
- Designed the LLM prompting layer to ground briefs in the actual data, not free-form summarization.
- Deployed end-to-end on a VPS with predictable cost and zero per-call infrastructure overhead beyond the LLM API.
- Shipped early access at spectre.asia for direct user feedback before broader rollout.
Result
- Live product running at spectre.asia, available for early access.
- Validated the core upload-to-brief flow end-to-end with real spreadsheets.
- Established the foundation for future productization (multi-tenant, billing, custom brief templates).
Tech stack
- Next.js
- Prisma
- SQLite
- OpenAI
- VPS Deployment