Product brief
Your email inbox, understood and acted upon — by an AI that works like a human assistant, not a rule engine.
Every professional's email inbox is a firehose of signal mixed with noise. Bills, receipts, shipping alerts, subscription renewals, flight changes, newsletters, bank statements — buried under promotional junk and notification spam. The information that matters is in there, but finding it, acting on it, and staying on top of it is a daily tax on your attention.
Today, the tools to manage this are either primitive (Gmail filters, Outlook rules) or require technical chops most people don't have (Zapier, n8n, custom scripts). Executive assistants and business managers solve this problem beautifully for C-suite leaders — triaging inboxes, tracking deadlines, surfacing what matters, filtering out what doesn't. But that service costs $60–120K/year and is available to a tiny fraction of professionals.
We're building an AI business manager for everyone. An always-on assistant that watches your email, understands it, and acts on your instructions — given in plain English, refined through conversation, and executed reliably in the background. Not a dashboard. Not a rule builder. A secretary you talk to.
Our customer is a working professional who gets dozens to hundreds of emails a day across one or more accounts. They're not a developer. They don't want to learn a new system. They don't want to maintain infrastructure. They want to describe an outcome in the same way they'd tell a human assistant, and have it happen.
Priya Sharma
Product manager, 32, Bengaluru
Has a personal Gmail, a work Outlook, and an HDFC bank account that sends transaction alerts by email. Subscribes to too many newsletters. Missed a credit card payment last month because the reminder was buried. Wants someone to just handle the boring stuff so she can focus on work and life.
Priya is representative, not exclusive. The product serves freelancers tracking invoices, small business owners monitoring expenses, parents managing household bills, anyone drowning in email who wishes they had a personal assistant.
The primary interface is a chat. The user talks to their secretary through whichever messaging channel they prefer — a purpose-built native app (with voice input and Action Button support), or through integrations with Slack, Telegram, WhatsApp, Signal, or a web UI. There is no settings panel, no rule builder, no dropdown menus. You describe what you want. The secretary asks clarifying questions if needed. Then it does it.
Every new rule gets validated against historical email before going live. The secretary runs a dry-run against past data and shows the user exactly what would have happened. If the output isn't quite right, the user refines it conversationally — "exclude transfers to my own accounts" or "separate online shopping from in-store" — and the secretary re-runs until it's right. At no point does it feel like programming.
Over time, the secretary observes patterns in the user's email and proposes automations. If it notices the user triages their inbox every morning on the commute, it offers a morning briefing. If it sees five Swiggy receipt emails a week that never get opened, it suggests filtering them. The system gets smarter the longer you use it — not because of complex ML, but because it accumulates context about what matters to you.
01
Conversation-first
Every interaction is a chat. No forms, no toggles, no rule builders. If you can't say it to a human assistant, you shouldn't have to say it to us.
02
Trust through transparency
Dry-runs against real data. The secretary shows its work. It says what it can't do. Trust is earned, not assumed.
03
Privacy as architecture
Open source. Self-hostable. E2E encrypted with customer-owned keys. We don't just promise privacy — the architecture enforces it.
04
Zero maintenance
No servers to manage, no rules to update, no infrastructure to worry about. Set it up once. It just runs.
The product is scoped to email as the primary input channel. This constraint is deliberate — email is universal, semi-structured, and rich with actionable data. By focusing on one input, we can go deep rather than wide.
Triage and filter. Classify incoming email by importance and category. Surface what matters, suppress what doesn't. Deliver a daily briefing of items that need attention.
Alert on conditions. Watch for specific emails and notify the user when criteria are met — bill amounts exceeding thresholds, delivery confirmations, flight schedule changes, subscription renewals approaching.
Track and accumulate. Log data from recurring emails over time — transaction amounts from bank alerts, spending by category, subscription costs. Deliver periodic summaries with breakdowns.
Remind and follow up. Set time-based reminders tied to email content — "remind me 2 days before the due date" or "flag this if I haven't replied by Thursday."
Answer questions about your email. "What am I paying for Figma?" "When does my insurance renew?" "How much did I spend on Amazon last month?" The secretary queries your email history and gives you a straight answer.
Draft replies. Compose responses to routine emails on your behalf, pending your approval. "Reply to the building manager confirming I'll attend the meeting."
Proactive insights. "You've been charged by a subscription you haven't opened emails from in 3 months." "Your AWS bill has increased 3 months in a row." "You have 12 newsletters you never open — want to unsubscribe?"
Cross-email correlation. "Your flight to Mumbai was rescheduled to 2:40 PM, but you have a 5 PM meeting — it'll be tight." Connect information across different emails to surface things a human assistant would notice.
The system is built in four layers. The bottom layer already exists as mailmux, an open-source, event-driven IMAP processing daemon written in Rust.
Interaction layer
Multi-channel frontend — native app (voice + text), Slack, Telegram, WhatsApp, Signal, web UI. A unified message bus normalizes input/output across all channels. The native app is the fastest path from thought to action; messaging integrations meet users where they already are.
Conversation engine
LLM-powered orchestrator that understands natural language instructions, manages multi-turn conversations, maintains context, and compiles user intent into executable rule specifications. Includes the dry-run engine that tests rules against historical email data for validation.
Rule engine and scheduler
Executes compiled rules deterministically — no LLM in the hot path. Event matchers evaluate trigger conditions against incoming email. A cron scheduler handles recurring tasks (weekly summaries, timed reminders). An accumulator maintains state across multiple emails for tracking and aggregation.
Email processing core (mailmux — exists)
Multi-account IMAP sync with IDLE push and polling fallback. Atomic event creation in PostgreSQL. Pluggable processor pipeline with retry and backoff. LLM-powered extraction of structured data from emails. Raw .eml storage. Replay and dry-run CLI. Prometheus metrics and health checks.
All three are adding AI to their email products. Gmail has Gemini summaries. Apple Intelligence does notification summaries. Outlook has Copilot. They will get better. They have insurmountable distribution advantages. So why build this?
Google's business model depends on reading your email. Apple's AI is locked to their ecosystem. Microsoft is enterprise-first. Our product is open source, self-hostable, and offers E2E encryption with customer-owned keys. A Google engineer cannot ship a product that says "we literally cannot read your data." We can.
Gmail's AI works best with Gmail. Outlook's Copilot works best with Outlook. Real people have messy email lives — a personal Gmail, a work Outlook, maybe a Fastmail or ProtonMail. We sit across all of them via IMAP, treating the email provider as a commodity pipe.
Big companies ship features for the median of a billion users — cautious, watered-down compromises. We can be opinionated. It's a secretary, not a platform. A conversational interface, not a settings panel. One thing, done with personality.
A Google PM will never greenlight an integration with a self-hosted personal finance app used by ten thousand people. But for the user who runs Firefly III, that integration is the entire reason to use our product. Open source lets the community build niche integrations that big companies will never bother with.
Apple and Google will send you notifications within their ecosystem. They won't let you talk to your secretary in Telegram or Signal or a custom Slack channel. We go where the user already is.
The project is fundamentally open source and completely self-hostable.
This isn't starting from zero. mailmux (github.com/avranju/mailmux) is an existing open-source Rust workspace that provides the email processing foundation: multi-account IMAP sync, atomic event creation in PostgreSQL, a pluggable processor pipeline with retry and backoff, replay and dry-run CLI, Prometheus metrics, Docker and systemd deployment, and an LLM-powered companion crate (mailtx) that extracts structured financial data from emails and posts to Firefly III.
The work ahead is building upward from this foundation — the rule engine, the conversation engine, the multi-channel interaction layer, and the native app.
The ones that matter will find her.