Public-tender startup · 2026
Scriva
I built the frontend and redesign of Scriva, an AI workspace that reads a tender, answers its questionnaire with cited sources, and drafts the bid with an AI co-writer.
Scriva is an AI bid-writing workspace for companies that answer government and commercial tenders. You upload the buyer's documents, and it reads them, lists what the buyer is asking for, answers the selection questionnaire from your own company files, and drafts the submission documents. Each bid moves through four stages, brief, compliance, draft and submit, until a finished PDF goes out.
It is hard to build because every AI step is slow and has to be trusted. Analysis and drafting run for minutes on background workers, so the interface has to stream honest progress and survive refreshes, double clicks and switching tabs mid-answer. And in a bid one wrong claim can cost the contract, so every answer has to point to the page it came from, and the AI can never quietly rewrite the document.
I owned the frontend from February 2026: the AI features, the document editor and its co-writer, the data layer and the contract with the backend, then a demo-readiness bug hunt and a full redesign of the product into Scriva's warm-paper system, built directly in code.
The decisions that made it work
Every answer points to a page
Questionnaire answers and Ask AI replies come back with citation markers. I parse them into chips naming the file and page, like ISO_9001_Certificate.pdf, page 1, that open the source PDF through a short-lived signed link, so any claim can be checked before it goes into a bid.
Long jobs you can trust
Reading a tender, answering a questionnaire and generating a bid take minutes. Their progress streams over SSE into a task tracker. Where the backend sends no percentage, the bar says Working rather than faking a number, and it never shows done before the job is.
A co-writer inside the editor
Drafts open in a Plate.js editor with an AI panel beside it. The agent streams its reasoning, keeps separate chat sessions, and sends edits as proposals. When a task finishes, the editor is replaced with the server's version so the two never drift.
Redesigned in code
I wrote the design brief, then rebuilt every screen to it in one pass: one clay accent on warm paper, a quiet icon rail, a calmer agent UI, from the bid desk to the auth pages, without changing how anything behaves.
The AI suggests, the writer approves
The agent never writes straight into the document. Each change arrives as a proposal the writer can apply or reject, one by one or all at once, and only then does the document change. In a bid, that control is the product.
4 stages
brief, compliance, draft, submit
Page-level
citations on AI answers
38K+
lines of frontend shipped
Every screen
redesigned in code