All workProDeck

An AI studio that turns a brief into client-ready work

ProDeck turns briefs, RFPs, meeting notes or a simple idea into proposals, presentations, documents and live microsites, written in your voice, styled in your brand, and fully editable.

Client
ProDeck
Year
2026
Duration
Ongoing
Services
Product design, AI engineering, Full-stack platform
A laptop on a desk running ProDeck, with the chat on the left and a generated yacht-builder proposal microsite open in the live panel
PII

Overview

ProDeck is an AI studio for client work. Each client gets a project workspace that holds their files, chat and memory. You describe what you need, ProDeck asks about anything missing, then builds the piece in a live panel beside the chat, grounded in your material and finished in your team's voice and brand.

Platform
Web app · TypeScript monorepo · CLI
Our role
Product strategy, UX and interface design, AI and agent engineering, Retrieval (RAG) pipeline, Document generation, Publishing and infrastructure
It produces
Proposals, Presentations, Documents, Live microsites
Headline
4 client deliverable types from one brief

The challenge

Client teams spend their best hours rebuilding the same proposals, decks and documents by hand. Generic AI tools are fast but invent facts, ignore the brand, and hand back output nobody can safely edit. 

The product

One workspace per client. A conversation on one side, the deliverable taking shape on the other.

The ProDeck workspace editing a client proposal PDF: a section title is selected in the live panel with a formatting toolbar, and the chat box on the left asks to describe the edit for that element
WorkspaceSelect any element in the live panel, then restyle it or describe the change in the chat.
  1. 01

    Bring the raw material

    Open a project workspace for the client and upload briefs, RFPs or meeting notes, or just start with an idea. The files and memory stay with that project.

  2. 02

    Describe what you need

    Ask for a proposal, a deck or a microsite in plain language. When something important is missing, ProDeck asks a clarifying question before it writes.

  3. 03

    Watch it build

    A multi-step planner breaks the request down and tool-using agents build the piece in a live panel, drawing on the uploaded material and the team's voice and brand.

  4. 04

    Refine by pointing

    Select any element and describe the change, optionally with a reference image. Undo, redo and version history keep every edit reversible.

  5. 05

    Ship it

    Publish microsites to a subdomain or your own domain, with optional password protection, or export documents to PDF, Word or PowerPoint.

  • Proposals
  • Presentations
  • Documents
  • Microsites
  • Your voice
  • Your brand

Capabilities

8 capabilities that take a team from a first brief to a published, on-brand deliverable.

  • 01

    Project workspaces

    Every client or piece of work gets its own space for files, chat and memory, so context never leaks between clients.

  • 02

    Chat-driven generation

    Describe the outcome and ProDeck asks clarifying questions, then builds the piece in a live panel next to the conversation.

  • 03

    Grounded output

    Content is retrieved from your own uploaded material, not invented, so claims in a proposal trace back to a source.

  • 04

    Author Voice

    Learns how your team writes and applies that voice to every proposal, deck and page ProDeck produces.

  • 05

    Design Kit

    Learns how your brand looks, from type and color to layout, and styles every output to match.

  • 06

    Click-to-edit microsites

    Select any element, describe the change and add a reference image if it helps. Undo and redo work across every edit.

  • 07

    Version safety

    Nothing is ever overwritten. Approved sections can be locked so later AI edits can't change them.

  • 08

    Publish and export

    Microsites go live on a subdomain or custom domain with optional passwords. Documents export to PDF, Word or PowerPoint.

Our approach

Three moves, shipped in weekly cycles, that turned generic AI into something a client team can put its name to.

  1. 01

    Start from the deliverable

    We worked back from what teams actually hand to clients, then designed one workspace where a chat on the left builds the piece in a live panel on the right.

  2. 02

    Ground every word

    Uploaded material is processed in Python and indexed with FAISS, so generation retrieves from the project's own sources instead of inventing facts. Zod validates every AI response before it renders.

  3. 03

    Make it safe to edit

    Click-to-edit microsites, versioning that never overwrites, and section locks mean the team stays in control of anything the AI touches, all the way to publish and export.

Architecture

A TypeScript monorepo with a plugin-based pipeline engine. Every request follows the same path, from raw input to validated, versioned output.

  1. 01 · Input

    • Briefs
    • RFPs
    • Meeting notes
    • Ideas

  2. 02 · Knowledge

    • Python parsing
    • FAISS index
    • Project memory

  3. 03 · Reasoning

    • Multi-step planner
    • Tool-based agents
    • Voice + brand

  4. 04 · Validation

    • Zod schemas
    • Versioning
    • Section locks

  5. 05 · Output

    • Microsite
    • PDF
    • Word
    • PowerPoint
Interface
  • Next.js 15
  • React 19
  • TypeScript
  • Zustand
  • TipTap
  • GSAP
  • Motion
API
  • Node.js
  • Fastify
  • Zod
Intelligence
  • Anthropic Claude
  • OpenAI
  • Ollama
  • Tool-based agents
  • Multi-step planner
Knowledge
  • FAISS vector search
  • RAG
  • Python document processing
Generation
  • Puppeteer
  • pdf-lib
  • docx
  • pptxgenjs
  • Mermaid
Platform
  • S3-compatible storage
  • pnpm workspaces
  • Plugin pipeline engine
  • CLI

Engineering decisions

AI that writes client work has to be reliable before it is impressive. These choices shaped everything else.

  1. 01

    One interface, any model

    Claude, OpenAI and local models through Ollama sit behind a single provider layer, so teams can choose on quality or cost, or keep sensitive work on their own hardware, without changing the product.

  2. 02

    Validate before it renders

    AI output is structured data, not free text. Zod schemas check every response, so malformed output is caught before it ever reaches the editor or a client.

  3. 03

    A pipeline built from plugins

    Ingestion, planning, generation and export are plugins on one pipeline engine. New output formats slot in without touching the core, and the same engine runs from the CLI.

  4. 04

    Versions, never overwrites

    Every AI edit creates a new version, and locked sections are off limits to the model. Trust comes from knowing nothing approved can silently change.

The results

0

deliverable types from one brief: proposals, decks, documents and microsites

AI providers behind one interface: Claude, OpenAI and local models
0
overwrites: every AI edit is a new version, and approved sections can be locked
0

Next project

A social app where time with friends pays you back

SocialApp · 2017–2021

Five SocialApp screens fanned across a deep blue field with gold Smart Coins: a Happy Diwali occasion thread, the Social Smart Card, the home screen at the center with Mood Talk, friends, news headlines, trending videos and stories, coins credited and a gift received, and the Redeem list of vouchers priced in coins

Bring us the brief

Let's turn your team's busywork into a product.

Start a projectAll work