Jums | Devpost

Inspiration

In today's world, we're drowning in a sea of hyper-specialized apps. We have apps for to-do lists, calendars, travel booking, food delivery, and communication. While each is powerful on its own, they exist in silos. The user is forced to become the "human API," manually shuttling information between them. This creates immense cognitive overhead and turns simple goals into a multi-app, multi-step chore.

Our inspiration came from a simple question: What if your planner could do more than just plan? What if it could act? We envisioned an AI companion that moves beyond passive task management to become a proactive orchestrator for your life, intelligently bridging the gaps between services to transform your goals into completed actions.

What it does

Jums is a proactive AI companion that functions as a "Life Orchestrator." Instead of you managing a list of tasks, you give Jums high-level goals in natural language, and it gets them done.

For example, you can say:

"Jums, plan a weekend getaway to Jaipur for me and my partner next month."

"Book a table for two at a good Italian restaurant near Connaught Place for Saturday at 8 PM."

"My flight lands at 3 PM tomorrow in Mumbai. Make sure I have a ride home and clear my afternoon schedule."

Jums' central reasoning engine deconstructs these goals into a series of logical steps. It then delegates these steps to a modular ecosystem of extensions, also called 'Jums'. These extensions are specialized plugins that interact with real-world services and APIs (like Google Calendar, MakeMyTrip, Zomato, or Uber) to execute each step. Finally, it presents a consolidated plan to the user for a simple, one-tap approval. Jums handles the complexity, you make the final call.

How we built it

We built Jums on a modern, microservices-based architecture designed for AI-driven workflows.

Frontend: A clean, conversational UI built with React Native for a seamless cross-platform experience on both iOS and Android. The focus is on minimizing clicks and allowing users to interact via natural language.

Backend: A series of Python microservices running on Google Cloud Platform. We chose Python for its robust AI/ML ecosystem.

The Orchestrator (The Brain): The core of our system is a custom-built reasoning engine. It uses a powerful Large Language Model (Google's Gemini API) for Natural Language Understanding (NLU) and task decomposition. This engine generates a Directed Acyclic Graph (DAG) of the plan, ensuring tasks are performed in the correct sequence.

'Jum' Extension Architecture: Each 'Jum' is an independent, containerized Python service with a standardized REST API. The Orchestrator communicates with these services to delegate tasks. For our prototype, we built three core Jums:

CalendarJum: Integrates with the Google Calendar API.

TravelJum: A wrapper around the Skyscanner and MakeMyTrip APIs for flight and hotel searches.

CommsJum: Uses Twilio to send notifications and confirmation messages.

Database: We used MongoDB to store user profiles, preferences, and the state of ongoing plans, leveraging its flexible schema to handle complex, evolving data.

Challenges we ran into

Ambiguity in Natural Language: A user's request like "find a nice place for dinner" is subjective. Training our orchestrator to ask clarifying questions instead of making poor assumptions was a major challenge.

API Chaos: Every third-party API has its own authentication, rate limits, and data schema. Creating a standardized "adapter" pattern for each 'Jum' to handle this diversity was a significant engineering effort.

Asynchronous Operations & State Management: Booking a flight isn't instantaneous. A plan might be in an "executing" state for minutes or even hours. Building a robust state machine to track the lifecycle of a goal from request to completion, including handling failures and timeouts, was incredibly complex.

Accomplishments that we're proud of

The End-to-End Workflow: We successfully executed a complete, non-trivial goal. A user was able to say, "Find me the earliest flight to Bengaluru this Friday and block out my travel time on the calendar," and Jums presented a real flight option and a calendar event for one-click confirmation.

The Modular 'Jum' System: Our plugin architecture works flawlessly. The orchestrator can dynamically discover and call upon the right 'Jum' for the job, proving the scalability of our ecosystem.

Hiding Complexity: We are incredibly proud of the user experience. Despite the immense complexity on the backend, the user's interaction is simple, elegant, and conversational. They see the result, not the process.

What we learned

LLMs as Planners: We learned that LLMs are more than just text generators; they can be used as powerful reasoning and planning engines to bootstrap agentic behaviour.

Human-in-the-Loop is Key: For a system with real-world consequences (like spending money), full autonomy is dangerous. We learned that the most critical feature is the final, user-friendly approval step, which builds trust and prevents errors.

Focus on the Orchestrator: The real intelligence of our system isn't in the individual 'Jums' (which are just API wrappers), but in the Orchestrator's ability to reason and direct them. This is where we will continue to invest our efforts.

What's next for Jums

Our vision is to make Jums the definitive OS for your life. The next steps are:

Expand the 'Jum' Marketplace: Open up our architecture to third-party developers to build and publish their own 'Jums'. We envision Jums for food delivery (Swiggy/Zomato), smart home devices (Google Home), finance (Zerodha), and team collaboration (Slack/Asana).

Proactive Assistance: Evolve from a reactive tool to a proactive one. Jums should be able to see a calendar event for a meeting across town and proactively ask, "This meeting is in 90 minutes and traffic is heavy. Should I book you an Uber?"

Deep Personalization & Memory: Implement a vector database to give Jums long-term memory. It should learn your preferences (e.g., you prefer window seats, you're a vegetarian, you never schedule meetings before 10 AM) to make smarter, more personalized decisions over time.

Launch an Invite-Only Beta: We aim to refine our prototype and launch a beta version for a small group of users in the next six months to gather crucial feedback.