Sourn

Inspiration

Anyone who buys supplies from more than one vendor knows the pain: the same item costs different amounts everywhere, pack sizes never line up, and the price you see is rarely the price you pay once shipping fees and order minimums show up at checkout. We wanted a tool that does all of that comparison for you and gives you a price you can actually trust.

What it does

Sourn finds the cheapest way to buy everything on your list. You give it your order, by typing or by talking to our AI voice assistant, and Sourn's agents compare suppliers, figure out the best way to split your order between them, and then verify the real checkout total. You get the cheapest verified plan, a breakdown of what to buy from each supplier, and how much you save from your original purchases.

How we built it

We built Sourn with a TypeScript and Express backend, a lightweight three-screen frontend, and Browser Use agents that interact with real supplier websites. A central orchestrator manages each comparison as a background job, tracks its progress, and coordinates the agents through discovery, optimization, and checkout verification.

The AI workflow has two stages:

Product discovery: Sourn launches one browser agent per supplier in parallel. Each agent receives the shopping list and strict instructions to remain on its assigned supplier’s website. It searches for each product, opens the relevant product pages, and returns structured data including the matched item, title, price, pack size, availability, SKU, and product URL. If an exact match cannot be confirmed, the agent must mark it as uncertain or missing instead of guessing.

Checkout verification: Once the backend creates the possible purchase plans, it determines which unique carts are required to evaluate them. These carts are grouped by supplier, allowing one checkout agent to test several cart combinations during the same browser session. For each scenario, the agent empties the previous cart, adds the correct products and quantities, and proceeds through guest checkout far enough to reveal the full cost breakdown. It records the subtotal, shipping, tax, fees, discounts, and final total, then saves a screenshot as supporting evidence.

The agents return schema-validated JSON rather than conversational text. Our backend treats this output as untrusted data and checks it before using it. It rejects off-domain links, unavailable products, uncertain matches, incorrect units, missing cart items, incorrect quantities, and incomplete checkout totals.

Between the two agent stages, deterministic code performs the actual optimization. It calculates how many packs are required to satisfy each requested quantity, chooses the best valid offer from each supplier, and generates every feasible way to assign the products across suppliers. After checkout verification, it sums the verified cart totals and selects the least expensive complete plan.

AI handles the unpredictable work of navigating websites, while regular code handles validation, arithmetic, and the final purchasing decision.

The frontend continuously polls the job state and displays the what the agent is doing live, allowing users to follow the search and checkout process before seeing the final verified plan.

CREAO

App ID: 0fc9749b-742b-459a-a5ec-6359937bd4b6 Link: https://agent.creao.ai/share/a86902f540a7ba1d8a9183012560a360

Challenges we ran into

Setting up the infrastructure. Most of us hadn't set up this kind of stack before, so getting the frontend, backend, and agents running together took real time and debugging.
Getting the AI voice assistant to work well. Integrating ElevenLabs was harder than we expected. We had to make the assistant respond quickly and stop it from repeating ** the same responses**, which took several rounds of fixes. We also had to speed up the process because we realize that talking to an AI robot in the first place is discouraging.
Listed prices aren't real prices. Shipping fees and minimums only appear at checkout, which is why we added the second pass of browser agents to verify actual totals.

Accomplishments that we're proud of

A full agent pipeline that goes from a shopping list to a verified checkout price, not just an estimate
A working voice assistant you can actually talk to that you can call and text
Normalizing products across suppliers so that different pack sizes compare fairly

What we learned

We learned how versatile these AI tools are. The same kinds of agents can browse websites, clean up messy product data, and hold a voice conversation. We also learned how much work goes into getting several AI systems to work together reliably, and that the hardest part of an AI product is often the infrastructure and details around the AI, not the AI itself.