Office

AI executive assistant that triages email by urgency, tracks commitments extracted from threads, generates tone-matched draft replies, and compiles meeting prep briefs. Runs locally with zero cloud exposure beyond the LLM API call.

What It Does

• Hybrid email classification. Rule engine scores 15+ signals (sender VIP status, direct vs. CC, urgency keywords, thread state, age, distribution size), then invokes Claude only for borderline cases (confidence < 0.7). Outputs P0-P3 priority with rationale. Most emails classify in <1ms. LLM handles ~15-25% of threads.
• Commitment tracking. LLM extracts structured commitments from P0/P1 threads: who, what, due date, direction (yours vs. theirs). Surfaces overdue items in the daily brief.
• Draft generation. Generates reply variants tone-matched to your writing style via a style learner that ingests your sent emails (rolling window of 50 examples).
• Meeting prep briefs. Before each meeting, compiles related email threads (by participant + subject overlap), open commitments involving attendees, and prior meeting context. Generates a 3-5 bullet prep brief.
• Daily command brief. Single view: urgent emails, today's meetings, overdue commitments.
• Full-text search. SQLite FTS5 across all emails, calendar events, and commitments.

Architecture

┌─────────────────────────────────────────────────┐
│          Dashboard (Next.js, port 3000)          │
│  Inbox · Thread detail · Calendar · Commitments │
└──────────────────────┬──────────────────────────┘
                       │ REST API
┌──────────────────────▼──────────────────────────┐
│           Server (Express, port 3456)            │
│  classifier.js ── hybrid rules + LLM triage     │
│  commitments.js ─ LLM extraction + tracking      │
│  meetingprep.js ─ context assembly + LLM brief   │
│  stylelearner.js  sent email ingestion           │
│  outbox.js ────── send via Power Automate        │
│  search.js ────── FTS5 across all entities       │
│  sync.js ──────── Graph API delta sync           │
│  db.js ────────── SQLite schema (11 tables)      │
│  llm.js ───────── Bedrock Claude wrapper         │
└──────────────────────┬──────────────────────────┘
        ┌──────────────┼──────────────┐
        ▼              ▼              ▼
   ┌─────────┐  ┌───────────┐  ┌──────────┐
   │ SQLite  │  │ MS Graph  │  │ Bedrock  │
   │ (local) │  │   API     │  │ Claude   │
   └─────────┘  └───────────┘  └──────────┘

Data flow: Power Automate exports emails to OneDrive drop folder, file watcher picks them up, Graph API delta sync for incremental updates, classifier scores, commitment extractor runs on P0/P1, everything stored in local SQLite.

Nothing leaves the machine except Microsoft Graph calls (email/calendar sync) and AWS Bedrock calls (LLM).

The Classifier

The hybrid classifier (classifier.js, 262 lines):

Rule-based first pass (deterministic, ~0ms):

• Sender VIP rules with wildcard matching
• Recipient position: direct To (+20) vs CC-only (-10)
• Urgency keywords: "ASAP", "blocker", "EOD" (+30) / "FYI", "no action" (-20)
• Ask detection: "please review", "can you", "approval needed" (+20-25)
• Thread state: awaiting your reply (+25), you replied last (-15)
• Age signals: unanswered >24h (+15), >48h (+30)
• Distribution: >20 recipients (-30), >50 (-50)

LLM second pass (borderline only, confidence < 0.7):

• Sends thread summary with rule signals
• LLM returns priority, needs-reply, and one-sentence rationale
• Email content marked UNTRUSTED with explicit instruction to ignore embedded overrides

Tech Stack

Layer	Technology
Frontend	Next.js, React, TypeScript
Backend	Express.js (ESM), Node.js
Database	SQLite via better-sqlite3 (11 tables + FTS5)
LLM	AWS Bedrock (Claude)
Email	Microsoft Graph API, Azure AD device code auth
Ingestion	Power Automate, OneDrive, local file watcher

API

Method	Endpoint	Description
GET	/api/brief	Daily command brief
GET	/api/threads	Prioritized inbox
GET	/api/thread/:id	Thread + messages + classification
POST	/api/draft/:id	Generate reply draft
POST	/api/sync	Trigger email/calendar sync
GET	/api/events	Calendar (7-day view)
GET	/api/commitments	Tracked follow-ups
GET	/api/search?q=	Full-text search

Quick Start

npm install && cd server && npm install && cd ../dashboard && npm install

cp server/.env.example server/.env
# Set: DROP_FOLDER, AWS credentials, USER_EMAIL

npm run dev  # API on :3456, Dashboard on :3000

Requires: Azure AD app registration (Mail.Read, Calendars.Read, User.Read) + Power Automate flow for email ingestion. See docs/POWER_AUTOMATE_SETUP.md.

Testing

55 tests covering security, reliability, and data integrity.

Run: cd server && node --test test/

License

Private