Easy Splitz

Inspiration

Splitting restaurant bills was always cumbersome. Doing split calculations manually—with Excel sheets or pen-and-paper methods, even using Splitwise—became exhausting. Manually confirming who ate what, entering each person's items individually, and managing payment disputes required constant back-and-forth. We envisioned a transparent, intuitive solution: snapping a photo of the bill, simply speaking who had what, and letting AI handle the rest.

What it is

• Upload & Scan: Users upload a bill image, which is automatically processed to capture items, prices, taxes, and tips.
• Easy of input: Casually tell who had what or how to split
• AI Calculation: Precise amounts owed by each participant is calculated with the help of AI tagging. Also, taxes and tips are automatically split proportionally.
• Transparency: Every participant can see itemized details explaining their charges.
• Real time iterative updating: Share the link to the split with friends, and everyone can watch or discuss the live updates as they get reflected in real time.

How we built it

1. Identified issues people face throughout the process
2. Defined MVP features for the prototype
3. Created detailed user flows and wireframes.
4. Developed and iteratively refined the prototype using Bolt.new.
5. Leveraging the advanced capabilities of Vision Transformer and GPT models, we utilized OpenAI to accurately extract comprehensive information from bill images, including individual items, prices, taxes, and tips. This significantly streamlined the data capture process.
6. We designed an intuitive, conversational interface allowing users to input splitting instructions casually and naturally. OpenAI then tags items to their respective consumers, followed by straightforward post-processing calculations. We incorporated rigorous validation checks to ensure each split accurately reflects the original bill.
7. Furthermore, we prioritized flexibility by enabling users to make incremental, conversational adjustments to splits. Users can easily add items or update individual contributions in real-time, with all participants able to view live updates, ensuring transparency and minimizing discrepancies. We refined our core features based on extensive feedback from diverse user experiences with receipt splitting.

Challenges we ran into

Balancing tradeoff between query flexibility (complexity of split instructions), accuracy, latency and token cost using different AI Split methodologies.

• Letting AI output direct amounts - low latency, low cost, high flexibility but not accurate as number tokens can be incorrect (Mental Math section in https://www.anthropic.com/research/tracing-thoughts-language-model)
• Letting AI output javascript code - high flexibility and extremely high accuracy (Research shows code output is more accurate than json output. Paper page - If LLM Is the Wizard, Then Code Is the Wand: A Survey on How Code Empowers Large Language Models to Serve as Intelligent Agents; huggingface/smolagents: 🤗 smolagents: a barebones library for agents that think in code. ) but high latency and high cost. We tried Cerebras which has much faster throughput but doesn’t yet support the latest best models which are more reliable to output code. Also vulnerable to malicious user attacks due to arbitrary code execution but that can be circumvented by using E2B.
• Letting AI tag items↔people who consumed what, followed by manual code calculation of split - medium flexibility (covers most of the use cases), high accuracy, low latency and low token cost - methodology we went ahead with

Accomplishments that we're proud of

• Successfully built a functional prototype delivering rapid and precise bill splitting.
• Easy iterations to splits supported
• Real time updates to a split
• Developed an intuitive user interface that significantly reduces manual entry.

What we learned

Throughout building Easy Splitz, we gained valuable insights into balancing trade-offs in AI-powered solutions. Specially learning how to use bolt.ai, how to vibe code & how to build from an idea to a live deployment with the least effort required. We learned that achieving the ideal balance between accuracy, cost, latency, and flexibility requires extensive experimentation.

Specifically, we found that:

• Allowing AI to output direct amounts provided low latency and cost but lacked accuracy due to token limitations.
• Using AI to generate executable code offered high accuracy but came at the cost of higher latency, complexity, and security risks.
• Opting for AI to tag items and associating them manually with users provided a practical balance of medium flexibility, high accuracy, and low latency.

This project reinforced our understanding of iterative development, emphasizing user experience and real-time updates as crucial elements for successful and intuitive solutions.

What's next for Easy Splitz

• Allow user to add a preferred payment method
• Allow users to pay their share in one click
• Allow user to easily correct when the split is wrong
• Real-time query support (“who had this item?”) with dynamic updates on everyone’s shares
• WhatsApp integration for effortless split calculation, bill sharing, disputes, and payments
• Mobile app development for broader accessibility
• Integration with Splitwise for seamless friend assignments and one-click export to user’s Splitwise

Built With

• bolt
• css
• firebase
• github
• lucide
• openai
• react
• tailwind
• typescript
• vite