Lofty AI Copilot

About Lofty AI Copilot

Inspiration

Real estate agents do not struggle with lack of tools. They struggle with too many disconnected tools and too many decisions.

A typical agent’s day involves figuring out which leads matter most, deciding who to follow up with and how, managing tasks appointments and deadlines, and manually tracking conversations across email and CRM.

We realized the real problem is not data. It is decision fatigue.

So we asked

What if the CRM did not just store information but actually told you what to do next and helped you do it

That question led to Lofty AI Copilot, an AI first workspace that acts like an intelligent partner instead of just a system of record.

How We Built It

We designed Lofty as a closed loop AI system that continuously understands the agent’s state, decides the next best actions, executes outreach, and learns from replies while updating context.

At a high level

DynamoDB People Properties Events Actions to App Router API app data to AI Assistant using OpenAI to execution layer with email and tasks to Gmail sync with OpenAI extraction to updated CRM state reflected in dashboard calendar and tasks

Key technologies used include Next.js App Router with TypeScript and Tailwind CSS for the frontend, OpenAI for structured decision making, ElevenLabs and browser speech recognition for voice, AWS DynamoDB for the database, AWS SES for email sending, Gmail API for reply ingestion, and Vercel for deployment.

We built several core systems including an AI assistant engine that converts natural language into structured actions and navigation, an action execution framework that turns recommendations into real outcomes like emails and follow ups, a Gmail reply intelligence layer that extracts intent such as rescheduling or preferences and updates the CRM, an AI powered calendar and task system that schedules work and detects conflicts, and a daily briefing with an Execute My Day flow that allows agents to run their entire day in seconds.

Challenges We Faced

One major challenge was making AI decisions deterministic. Language models are flexible but product workflows require structure. We solved this by designing typed assistant responses where the AI outputs clear intent targets and structured payloads.

Another challenge was closing the loop between execution feedback and system state. Most AI demos stop at generating content, but we built a full loop where emails are sent through AWS SES, replies are received through Gmail API, structured meaning is extracted using OpenAI, data is updated in DynamoDB, and the UI reflects changes immediately.

We also had to design for both demo and real world usage. We balanced real integrations with hackathon constraints by converting calls and SMS into email safe flows, adding undo states for repeatable demos, and including fallback mock data.

Finally, building a context aware assistant required aggregating information from the current page selected lead recent conversations and pending actions so the assistant could handle follow up queries naturally.

What We Learned

AI is most powerful when it can take action instead of only generating responses. The real user experience breakthrough is decision automation. Closing the loop between communication understanding and system state creates strong leverage. Even with powerful models structure and constraints are necessary for reliability. The best AI products feel less like tools and more like operators.

What Makes This Different

Lofty AI Copilot is not just a CRM chatbot or automation tool. It is a system that understands your pipeline decides what matters executes work and keeps everything in sync automatically.

Built With

• aws-dynamodb
• aws-ses
• browser-speechrecognition
• elevenlabs-text-to-speech
• gmail-api
• lucide-react
• next.js-app-router
• openai-api
• shadcn-style-ui-components
• tailwind-css
• typescript