Proactive AI

An approachable, proactive AI agent that knows you better than you do

Comment

Inspiration

We noticed that businesses are drowning in context switching. Employees spend hours jumping between Slack, email, calendars, and productivity tools just to get basic tasks done. Important invoices get buried. Meeting requests sit unanswered. Flight changes get missed. We realized the problem isn't that people need better tools. They need tools that actually understand them and take action proactively. So we built this.

What it does

A proactive AI agent that lives in iMessage and actually knows you. It connects to your email, calendar, and daily context to anticipate what you need before you ask. Instead of waiting for prompts, it texts you throughout the day like a coworker who actually pays attention:

  • Surfaces important emails with one-tap actions (invoice due Friday? tap to pay)
  • Remembers every detail you mention and uses it weeks later (you mentioned liking matcha? here's that new spot)
  • Journals with you nightly to understand your schedule and preferences
  • Handles meeting requests, travel changes, and payment reminders without app switching
  • Builds genuine memory that feels like texting a friend who cares

For businesses, this means employees spend less time managing their digital lives and more time on actual work. It reduces context switching, catches missed tasks, and automates the boring stuff through simple text conversations.

How we built it

We built it on a custom agentic architecture inspired by Claude Code. The backend runs on Cloudflare Workers with a real-time WebSocket architecture for instant messaging. We use Docker containers running Ubuntu Linux for sandboxed code execution, letting it safely run tasks and generate files. Our memory system combines PostgreSQL for structured data with Qdrant vector database and GraphRAG for semantic search. We integrated 135+ tools and libraries so it can take real actions across different platforms. The AI layer uses Claude Sonnet 4.5 for reasoning and Haiku for speed. We built the iOS interface in SwiftUI and added a voice interface powered by Cartesia. Email and calendar integrations use native APIs. The agent has automatic retry logic with dynamic model selection. Everything from voice input to task execution to file delivery works end to end. We also built browser automation with Playwright so it can actually do things on the web.

Challenges we ran into

Building proactive AI is way harder than reactive AI. The biggest challenge was making it feel alive without being annoying. We had to figure out variable response timing so it doesn't feel robotic. We struggled with the memory system because remembering everything creates noise. Deciding what to remember and when to surface it took constant iteration. The agentic loop was complex because we needed it to reason about tasks, execute them safely in Docker, handle failures gracefully, and deliver results naturally through text. Getting the voice interface to work smoothly with Claude and code execution was painful. We also had to solve real-time WebSocket communication at scale on Cloudflare Workers. The email parsing was rough because every inbox is chaotic. Training ourselves to think like the product by actually texting people constantly helped us understand what retention actually feels like.

Accomplishments that we're proud of

We built a genuinely proactive AI that doesn't wait for prompts. It actually texts you first with things you need. The memory system works incredibly well. It remembers tiny details from weeks ago and surfaces them at exactly the right moment. Our nightly journaling feature creates genuine rituals that users look forward to. The one-tap actions for email tasks actually work. Invoice? Paid. Meeting? Rescheduled. Flight? Rebooked. All without leaving iMessage. We're proud of the technical architecture too. The sandboxed execution environment lets it safely run code and generate files. The voice interface is seamless. The agent completes end to end tasks from understanding what you need to executing it to delivering results. Most of all, we're proud that it genuinely feels different. It doesn't feel like a chatbot. It feels like texting someone who knows you.

What we learned

We learned that retention mechanics are predictable. Memory builds trust faster than anything. Status updates drive engagement. Nightly check-ins create rituals. Variable response timing makes things feel alive. Being too available kills engagement. We also learned a ton technically. Building agentic systems that can reason, execute, and recover from failures is hard. Working with vector databases and semantic search for memory retrieval taught us how to make AI actually remember things that matter. Integrating 135+ tools and making them work together smoothly was a crash course in API design. We learned that proactive AI requires completely different architecture than reactive AI. You can't just wait for inputs. You need to monitor context, decide when to interrupt, and take action confidently. On the less technical side, we learned to move fast under pressure and ship something that actually works in 24 hours.

What's next for Proactive AI

We're taking this to market. Businesses need this more than anyone. Every company has employees drowning in email, missing deadlines, and context switching constantly. It eliminates that friction. We're expanding integrations to Slack, Microsoft Teams, and enterprise tools. We're building team features so it can coordinate across people and surface insights to managers. We're adding commerce capabilities so it can actually buy things for you when prices drop or remind you about subscriptions. We're training the memory system on more data to get even better at predicting what you need. Long term, we want it to be the interface layer for your entire digital life. Not just an assistant, but a friend that actually knows you and handles everything proactively. We're moving to SF after this to build it properly.

Built With

Share this project:

Updates