AssistantAI

An AI-powered personal assistant. Speak (or type) a messy brain dump of everything you need to do today, and AssistantAI extracts every task and fixed event, estimates missing durations, prioritizes, detects conflicts, and builds one optimized daily schedule with a plain-language explanation of its reasoning.

"I have class from 9 to 11. I need to finish my SQL assignment which
should take about two hours. I have work from 3 until 11. I should go to
the gym before work. I need groceries. I should call my mom. I need to
study Python for an hour."

becomes a full, conflict-checked timeline — class, gym, the assignment, groceries, study time, lunch, work — in one shot, with a note on why it was ordered that way.

Live demo: https://assistantai-sooty.vercel.app

Tech stack

• React + Vite, React Router
• Bootstrap 5 (+ plain CSS) — no Tailwind
• Supabase (Postgres + Auth + Row-Level Security)
• Vercel serverless functions
• Anthropic API (Claude), server-side only

Architecture overview

Browser (React SPA)
  |
  |-- Supabase Auth (email/password) --> Supabase (Postgres, RLS)
  |
  |-- POST /api/plan (Bearer <user JWT>)
        |
        v
  Vercel serverless function (api/plan.js)
    1. Verify JWT against Supabase, scope a Supabase client to that user
    2. Call Claude with a forced tool call (submit_daily_plan) that:
         - extracts fixed events + flexible tasks
         - estimates missing durations, assigns priority
         - builds one non-overlapping optimized schedule
         - writes a short explanation + any conflicts
    3. Re-validate the structured output before writing anything
    4. Upsert daily_plans, replace its tasks + calendar_events,
       log the call to ai_sessions
    5. Return the full assembled plan to the browser

The frontend never talks to Anthropic directly and never sees the API key. Every database table is protected by Postgres RLS policies scoped to auth.uid(), so the anon key is safe to ship in the browser bundle.

Database schema

All tables live in supabase/schema.sql and are RLS-protected (auth.uid() = user_id).

Table	Purpose
profiles	One row per user. preferences jsonb holds forward-compatible scheduling defaults (wake/sleep time, focus block length) — not read by any code yet.
daily_plans	One brain-dump + generated schedule per (user_id, plan_date). Holds the raw transcript, the AI's explanation, detected conflicts, and which model produced it.
tasks	Flexible to-dos the scheduler placed: category, priority, duration, deadline, depends_on_task_id, status (pending/scheduled/completed/cancelled), and scheduled_start/scheduled_end.
calendar_events	Fixed blocks with a hard start/end (class, work, appointment). external_provider/external_event_id are reserved for future Google/Apple Calendar sync — unused today.
ai_sessions	Audit log of every /api/plan call: request payload, raw Claude response, status, latency. Used for debugging and future habit-learning.
integrations	Not implemented. Placeholder for future Google/Apple Calendar OAuth connections so that feature won't need a schema migration later.