ActAs.Me (Email, sorted.)

Inspiration

Email is a 50+ year old invention, but attention has never been scarcer. Executives spend hours a day triaging inboxes full of cold pitches, scheduling back and forth, and low value noise, when what they really need is a handful of decisions that actually matter. We wanted to build the last inbox an executive would ever have to manage themselves. That is where actas.me came from: an autonomous executive assistant that acts as you, so you do not have to.

What it does

actas.me sits in front of your inbox as a proactive agent, not just a chatbot. It ingests incoming email in real time, classifies intent, and filters noise. It then plans and executes multi-step actions, like checking calendar availability, proposing conflict free meeting times, and syncing updates to your CRM, all without manual intervention. When a decision genuinely needs a human, actas.me surfaces a single one tap approval card on your phone. Approve it, and actas.me sends the authenticated reply and updates your systems on your behalf.

How we built it

• Google Gemini 1.5 Flash handles high speed, low cost initial email classification, input sanitization, and structured JSON parsing as messages arrive.
• Google Gemini 1.5 Pro handles multi step reasoning, contextual document understanding for attachments like contracts and PDFs, persona matching, and tool planning.
• Google Text Embeddings power semantic memory retrieval over past conversations and tone examples, so actas.me learns how you actually write and negotiate.
• Model Context Protocol (MCP) standardizes tool calling so Gemini Pro can invoke external tools, starting with Google Calendar, cleanly and predictably.
• Cloudflare Workers, Durable Objects, Vectorize, and R2 provide the serverless edge execution pipeline, from receiving the inbound email to storing state and vector memory.

Challenges we faced

Getting an agent to act autonomously without acting recklessly was the hardest part. We had to build guardrails against prompt injection hidden inside incoming emails, add a Shadow Mode so the agent's proposed actions could be reviewed before going live, and carefully scope which actions could run automatically versus which always required the one tap approval step. Chaining Gemini Flash and Gemini Pro together, so classification stays fast while reasoning stays deep, also took real tuning to get the handoff right.

Accomplishments that we're proud of

We got a full end to end loop working: live email ingestion, classification, autonomous tool calling against Google Calendar, and a mobile approval card that triggers a real authenticated send and CRM update. It genuinely feels like handing off your inbox to someone you trust.

What we learned

Autonomy is not about removing the human, it is about being precise about which moments actually need a human. Most of an executive's inbox does not need a decision at all, and building that judgment into the agent mattered more than raw model capability.

What's next for actas.me

Expanding beyond calendar and CRM into fuller negotiation flows, deeper long term memory of relationships and tone, and broadening tool support so actas.me can act as you across more of your daily operational surface area.

Built With

• cloudflare
• durable-objects
• gcp
• gemini-flash-2.5
• gemini-pro-2.5
• mcp
• pubsub
• vectorise
• vertixai