What if a leadership team’s priorities, scorecards and issues could keep themselves current?

A business operating system for senior-care leadership teams: vision, accountability chart, weekly scorecards, priorities, issues, processes and a knowledge base — with Olga, an assistant that answers from the organisation’s own data and drafts entries conversationally. I built it from the first commit to production.

What I did

Set up a Turborepo monorepo — Next.js web app on Vercel, NestJS API on Railway, shared packages for contracts, database, UI and config — with GitHub Actions running lint as a required gate and path-filtered deploys on push.

Built multi-tenant isolation on PostgreSQL row-level security: the API runs as a restricted database role and binds the tenant on every request, and org isolation is proven against a real Postgres in CI on every pull request.

Implemented identity with Better Auth — the web app mints JWTs, the API verifies them through JWKS, and a BFF proxy lets the UI call the API as the signed-in user — plus email verification and password reset through Resend, failed sign-in auditing, security headers and per-user throttling.

Shipped the product surfaces: Vision Board, an interactive Accountability Chart (React Flow + dagre), a weekly Scorecard with per-seat measurables and a past-week lock, Priorities → tasks → a planner week view, an Issues board with threaded discussion, a versioned Processes library, audience-targeted Announcements with read receipts, and an S3-backed Knowledge Base with generated covers.

Built Olga on the Vercel AI SDK and Anthropic: streaming chat docked on every screen, read tools grounded in RBAC-scoped org data, priority-to-task breakdown proposals the user approves before anything is written, persisted conversations, and a superadmin transcript review with provenance.

Wrote the product docs from the PRD, deterministic seeds for every seat in the accountability chart, and kept the stack current through TypeScript 6 and Next.js 16 upgrades.

Architecture

Two deployables and a handful of shared packages, managed with Turborepo and pnpm workspaces.

apps/web — Next.js 16 on Vercel. Server components for the product screens, route handlers for the BFF proxy and the AI endpoints, Better Auth for sessions.
apps/api — NestJS on Railway with a managed PostgreSQL. Every request runs as the app_runtime database role with the tenant bound through a per-request interceptor, so row-level security — not application code — is what keeps organisations apart.
packages/db — Drizzle ORM schema, migrations and the RLS policies, with isolation tests that run against a real Postgres in CI.
packages/api, ui, eslint-config, typescript-config — shared contracts, the component library and toolchain presets.

Identity

The web app mints short-lived JWTs; the API verifies them through a JWKS endpoint and maps the subject to a database role and tenant. The UI never holds an API token — a BFF proxy forwards calls as the signed-in user. Auth events (sign-in, failures, resets) are audited at the enforcement point.

Olga

Olga is a streaming chat panel available on every dashboard screen, built on the Vercel AI SDK with Anthropic models. Her tools are read-only and scoped by the caller’s RBAC permissions, so she can answer “what slipped this week?” from scorecards, priorities and issues without seeing anything the user could not. Writes — a priority broken into tasks, an issue drafted from a conversation — are proposed, shown, and only committed after the user approves. Conversations are persisted, and a superadmin review surface shows transcripts with provenance for every tool call.

What I would do next

Phase 2 of the PRD: proactive nudges from Olga (a seat with no accountabilities, a measurable missed three weeks running), and the meeting agenda that assembles itself from the week’s data.

RunOlga

What I did

Built with

Screens

Architecture

Identity

Olga

What I would do next