Starling MX
A live proof of concept became an enterprise AI knowledge platform in six months, without going offline.
Free 30-minute call. You leave with a written summary of the approach and a sensible first step. No deck, no pitch.
- Billed hours
- 1,145
- Duration
- 6 months
- Client rating on Upwork
- 5.0
Stack
- Next.js 15
- Supabase RLS
- Claude API
- OpenAI API
- Notion API
Removed line: Email the access request form to IT.
Added line: Ask a workspace admin to approve you in Settings.
Written to knowledge base
Edit dismissed. The knowledge base is unchanged.
The starting point
The client had a live proof of concept: an AI knowledge platform that already answered questions from the company's documents and already had people using it. To be sold into enterprises it had to do four things the prototype could not. Process real PDFs, Word files and slide decks without freezing whoever uploaded them. Keep every workspace's documents separate from every other's in a way an auditor could verify. Keep the model bill predictable as conversations grew long and documents grew large. And stop the AI from changing the knowledge base without a person agreeing to it.
What we built
01
Background document conversion
PDF, DOCX and PPTX files convert in background jobs, one converter per format, so a large upload never holds up the person using the app and a failed conversion retries on its own.
02
Prompt caching and a context budget
Prompt caching on the stable part of each prompt, and a per-request context-window budget split between history, retrieved passages and the answer, so long conversations stay within limits and within budget.
03
Streaming chat in multi-tenant workspaces
Streaming AI chat over Claude and OpenAI runs inside multi-tenant workspaces, with Supabase row-level security keeping each tenant's data separate at the database rather than in application code.
04
Human approval before AI edits
When the AI proposes a change to the knowledge base, a person sees the proposed edit with the evidence behind it and approves or rejects it before anything is written back.
What it came to
An enterprise-ready product built on the original proof of concept, which stayed live throughout. 1,145 billed hours over six months, rated 5.0 by the client. The review that came with the rating is quoted on the homepage: a vague problem goes in and a scoped plan with the trade-offs named comes back.
Have a proof of concept that enterprise buyers are asking hard questions about? Book a free architecture call and bring the questions. The discovery sprint that follows is the audit, from $2,500.
Book a free architecture call