entri

AI study-notes app: photograph handwritten notes → extraction that never silently corrects (suspect values are surfaced as tentative corrections, not rewritten) → an FSRS daily review tuned to your exam date → grounded RAG chat over your own notes. Thesis: retention + trust.

• Product/eng plan and milestones: PLAN.md
• Design system (read before any UI work): DESIGN.md

Monorepo

This is a Turborepo managed with pnpm workspaces.

apps/
  api/         Hono API + in-process ingest worker (Node, port 8787)
  web/         Next.js app (port 3000)
packages/      shared code (none yet)
migrations/    InsForge SQL migrations
insforge.toml  InsForge project config
PLAN.md        CEO/eng plan, milestones M0–M6
DESIGN.md      Design system (Marginalia, dialed back)
turbo.json     task pipeline   ·   pnpm-workspace.yaml

Layer	Tech
Web (@entri/web)	Next.js (React) — apps/web, port 3000
API (@entri/api)	Hono on Node (@hono/node-server) — apps/api, port 8787
Backend (DB, auth, storage)	InsForge (Postgres + pgvector)
Extraction	Sarvam Vision OCR → card structuring on DeepSeek V4 Flash (OpenRouter)
RAG chat + concept graph	DeepSeek V4 Flash via OpenRouter (InsForge AI gateway)
Embeddings	Cloudflare Workers AI (bge-large, 1024-dim)

The apps/api ingest worker is an in-process queue (polls the jobs table). It's slated to become a Cloudflare Workflow/Queue later (see apps/api/src/worker.ts).

Prerequisites

• Node 20+ and pnpm (npm install -g pnpm)
• An InsForge project (free) — for DB/auth/storage
• (Optional, for capture/chat) an OpenRouter API key (the InsForge AI gateway, for the chat/text model), a Cloudflare account with a Workers AI API token (embeddings), and a Sarvam API key (OCR). Without these the app still runs: auth, UI, and browsing work; only photo capture (extraction) and chat are disabled.

Setup

# 1. Install the whole monorepo (one command)
pnpm install

# 2. Configure environment
cp apps/api/.env.example apps/api/.env        # InsForge + (optional) Cloudflare/Sarvam keys
cp apps/web/.env.example apps/web/.env.local  # NEXT_PUBLIC_INSFORGE_* (anon key)

# 3. Set up the database (applies migrations/ to your InsForge project)
npx @insforge/cli link
npx @insforge/cli db migrations up --all

# 4. Run everything (Turbo runs api + web in parallel)
pnpm dev

• Web: http://localhost:3000 · API: http://localhost:8787

Open the web app, create an account, and you're in.

Optional: seed a demo deck

# Find your user id from GET /v1/me after signing up, then:
pnpm seed -- <your-user-id>

What runs without the AI keys

apps/api/.env gates the AI stack on OPENROUTER_API_KEY (chat/text) + CLOUDFLARE_* (embeddings) + SARVAM_API_KEY (OCR). If those are unset:

• ✅ Auth, dashboard, notes browsing, settings, review of seeded cards — all work.
• ⚠️ POST /v1/capture returns 503 AI not configured; the ingest worker logs idle.
• ⚠️ /chat returns 503 (needs OpenRouter for the chat model + Cloudflare for embeddings).

So you can boot the full UI with just an InsForge project, and add the AI keys later.

Commands

Run from the repo root (Turbo fans out to the workspaces that define each task):

Command	Does
pnpm dev	api (watch) + web dev server, in parallel
pnpm build	production build (web)
pnpm lint	ESLint across workspaces
pnpm typecheck	tsc --noEmit across workspaces
pnpm seed -- <userId>	load the demo deck for a user

Target one workspace with a filter, e.g. pnpm --filter @entri/web dev or pnpm --filter @entri/api typecheck.