An AI-streaming rich-text editor for an EU tendering firm. Plate.js for the canvas, Claude for the inline completions, a custom DOCX exporter for the deliverable. A separate Python AI service handles the heavy retrieval work in the background while the editor stays at 60fps.
Proposal writers in the public-tender world live in two windows simultaneously — a Word document and a research browser. Scriva collapses those into one. A Plate.js editor with Claude streaming directly into the canvas, source-cited paragraphs anchored to the originating document, and a one-click export that produces a DOCX a procurement officer cannot tell from one a human typed. Behind the editor sits a Python AI service that handles the heavy retrieval and a document service that handles parsing and storage — three apps in a monorepo, each independently deployable.
EU public tenders are won and lost in the editor. A tendering team typically receives a 60-to-200-page request-for-proposals document, has 14 to 28 days to respond, and produces a deliverable that must conform to the original document's structure, terminology, and citation expectations *exactly*. Off-by-one section numbering disqualifies. Mis-stating a regulatory reference disqualifies. Using a phrasing that doesn't track the source-language definition can disqualify.
So writers live in two windows. Word, where the proposal is being typed. Browser, where the RFP PDF and a dozen regulatory references are open. They alt-tab thousands of times per proposal. They paraphrase by hand. They lose citations between paragraphs. And the final DOCX must be assembled by hand into the template the procuring authority distributed — including font, margins, header logo, signature block.
“Writers were alt-tabbing thousands of times per proposal and rebuilding the DOCX template by hand at the end of every cycle. The AI tools that existed didn't fit the actual job.”
Generic AI editors are nowhere near this work. Notion AI is built for blog posts. ChatGPT lives outside the document. Microsoft Copilot is Word-shaped and not great at multi-document retrieval. The team had tried all three and was building proposals the old way regardless.
What they wanted was specific: a real rich-text editor with the formatting fidelity to produce a regulator-grade DOCX, an AI that could be invoked *inside* a paragraph and stream tokens into the live document, citations that linked to the source PDF, and a retrieval layer that already knew which regulations applied to which proposal type.
AI completions stream *into* the paragraph, not into a side panel. The writer keeps their hands on the document.
Most teams that quoted them an AI editor planned to wrap an existing editor library and bolt a side panel onto it. Side panels are the wrong product. Writers don't want to read AI output in a different surface and copy-paste it across — they want the AI to type *into* the paragraph they're already in. That's a hard editor problem, not an AI problem, and the team that solved it had to be comfortable extracting and modifying a real rich-text engine.
We chose to base the editor on Plate.js (which sits on Slate) precisely because it gives us extension points all the way down to the rendering loop. We pulled the relevant portions of Plate into our codebase via an internal registry pattern — editor self-contained, zero hidden runtime dependencies except `@/lib/utils`. That decision unlocked everything downstream: streaming AI directly into a paragraph, custom decorations for citation links, DOCX export with full formatting fidelity, all on a single in-house editor surface.
Place the cursor anywhere, invoke the AI with a hotkey or floating toolbar, and Claude tokens stream *into* the paragraph in real time. Subtle teal highlighting marks AI-authored ranges so the writer can review and lock down language before export.
Every AI completion is anchored to the originating document and page range. Click a citation marker in the editor to open the source PDF at the cited paragraph. Citations survive DOCX export as footnote-style references.
Single-click export to a DOCX that preserves headings, lists, tables, footnotes, page breaks, and embedded images. Built on `@platejs/docx-io` with custom serializers for the structures procurement authorities expect.
The AI service indexes the active RFP plus a configurable corpus of regulatory and prior-proposal references. Completions and rewrites are grounded against this corpus and never fabricate a citation that isn't in the indexed set.
The AI toolbar adapts to the current section's type — executive summary, technical approach, pricing schedule, references — and offers context-appropriate actions: Rewrite formally, Compress, Expand with citation, Convert to bullet list.
Every proposal is versioned at the section level. Writers can compare two versions side by side, restore a previous section, and audit every AI invocation by author, prompt, and timestamp.
Single-click export to a DOCX a procurement officer cannot tell from one a human typed.
MethodologyPerformance numbers measured on the production staging environment with a representative proposal corpus. Export fidelity verified against three procurement-authority DOCX templates the client routinely receives. Versioning audit-grade is a property of the data model.
The first AI editor where I stop noticing the AI. I write, it offers, I keep writing. The DOCX at the end is what we used to spend a Saturday assembling.
Multi-language proposal mode — English, German, French side by side with linked sections. A team-collaboration layer for parallel writers on the same proposal. Deeper retrieval against the firm's historical proposal corpus so completions can cite the firm's *own* prior winning language.
Generic AI editors don't survive contact with regulator-grade documents. If your team produces tenders, legal filings, clinical reports, or any deliverable with strict format and citation requirements — we can scope an in-house editor that meets the format, not just the language.