Case study: RPOS - a working prototype of an agent-assisted PM workspace

Personal project, 2026. Conceived, designed, and built end to end by me, with AI pair-assistance. Prototype stage - honest scope notes below.

The premise

A product manager's artifacts - personas, user stories, backlog, story map, PRD - live in five disconnected tools and go stale the moment they are written. Change a persona and the stories quietly stop matching it; reprioritise the backlog and the PRD nobody re-exported still says otherwise. RPOS (Rutu's Product OS) prototypes the opposite idea: one shared workspace state, where every PM artifact is a live view of it, and the PRD regenerates on every keystroke.

The full concept is a 4-agent pipeline that takes a raw idea through Vision, Research, PRD/Backlog/Story Map, and UI mocks. This prototype pass makes Agent 3 - the core PM workspace - fully functional, and visualises the rest.

What is actually built

• The working prototype (single HTML file, React 18, no build step): a three-pane workspace - SDLC sidebar, active module, live PRD preview. Five editable modules: Strategy (vision, value prop, goals, North Star), Personas (six empathy maps), Stories (Given/When/Then acceptance criteria), Backlog (4-column Kanban with drag and drop and MoSCoW tags), and Story Map (activity-by-release swimlanes). The PRD pane regenerates from shared state live and exports to .md or .doc. A KPI strip computes coverage stats from state.
• A 12-screen visual mock of the full product in a Figma-style shell - dashboard, roadmap with a 26-week Gantt, export modal, the 4-agent pipeline view.
• A research-tool front end (React + TypeScript) for the research agent: claim-by-claim output with sources panel and follow-up flow.

Design choices

Single source of truth over integrations. The standard answer to artifact drift is syncing between tools. RPOS bets the other way: one state object, many renderings. That is why the live PRD is the centrepiece - it is proof the state is coherent.

Discipline baked in, not bolted on. Empathy maps, Given/When/Then, MoSCoW, story mapping with release slices, INVEST-style sizing - the frameworks are the data model, so the tool makes the craft the default rather than a template you paste.

Zero install. The prototype is one HTML file that runs from a double click. A PM tool you can demo anywhere, including in an interview.

Honest scope

This is a prototype, not a product. One of four agents is functional; research and UI-mock agents are visualised. No persistence, no auth, no users to cite. What it demonstrates is judgment and build ability: scoping a credible V1 slice, shipping it working, and being precise about what is real.

Next iterations

Wire the prompt and research agents to a live LLM backend, persist workspace state, and replace the static UI-mock agent with a live design-tool integration.

Ruturaj Sahasrabudhe | ruturaj.mms@gmail.com | linkedin.com/in/ruturajsahasrabudhe