Mercora

Inspiration

Small startups or businesses that deal with perishable supply often know exactly when something is going wrong, a tray of pastries may go unsold before close, appointment slots may remain empty next week, or expensive inventory may sit on the floor, but acting quickly still requires several disconnected kinds of work. The owner must understand the problem itself, create marketing, update the website, understand customer feedback, and decide whether to produce, discount, reorder, or stop promoting. Most tools solve only one piece of that problem.

What it does

That insight became Mercora, an AI growth operator that begins with a verified business need, creates a truthful owner-approved action system, measure the resulting evidence, and recommends the next operating decision. Instead of stopping at website generation, content creation, analytics, or forecasting, Mercora connects them into one continuous feedback loop and consistently gets better.

How we built it

We built Mercora as two tightly connected systems: an owner-facing AI workspace and a deterministic business-decision engine. The AI handles conversation, workflow planning, content, and explanations; dedicated numerical services handle forecasting, causal-lift estimation, simulation, and optimization. This separation lets Mercora feel agentic without allowing a language model to invent demand forecasts, savings estimates, or operating decisions. We used GPT-5.6 Codex Goals to build Mercora in parallel from a shared base. Before implementation began, we froze the core OpenAPI contract, identifiers, fixtures, and domain semantics so separate agents could work independently without inventing incompatible data models.

Challenges we ran into

The hardest challenge was making uncertainty and business risk easy for owners to understand. Mercora needed to show ranges, assumptions, and evidence quality without turning the product into a statistics dashboard. We also had to clearly separate views, engagement, attributed conversions, estimated lift, and actual outcomes. Another challenge was keeping GPT-5.6 useful without letting it invent decisions. The model handles conversation, planning, content, and explanations, while dedicated forecasting, simulation, and optimization services produce the numerical results. Building the engine and frontend in parallel created integration issues around CORS, request formats, state sequencing, asset loading, and environment differences. We solved this by freezing shared contracts and testing the real browser and backend together. Approval and publishing also required careful design. Mercora now binds approval to the exact campaign payload, invalidates approval after material changes, prevents duplicate publication, and reconciles uncertain provider states safely. Finally, we had to make the whole system reproducible across local development, clean installs, browser tests, and GitHub Actions. The biggest lesson was that a strong AI product depends not only on features, but also on trustworthy data, safe actions, consistent state, and reliable recovery when something fails.

Accomplishments that we're proud of

We are most proud of building Mercora as a complete operating loop, not just another AI content tool. It can move from a verified business need to a campaign, approval, publication, customer signals, updated forecast, and cost-aware recommendation. We also built strong trust and safety into the product. Mercora separates engagement from conversions and causal lift, binds approval to the exact campaign payload, prevents duplicate publication, shows uncertainty, and can abstain when evidence is weak. Technically, we created one configurable system that works across perishable inventory, appointment capacity, and slow-moving retail products. The AI handles conversation, planning, content, and explanation, while dedicated forecasting, simulation, and optimization tools produce the numerical results. Finally, we are proud of the reliability of the system. Mercora includes tenant isolation, audit trails, deterministic testing, browser-to-backend verification, responsive design, and recovery from ambiguous provider failures.

What we learned

We learned that the most useful AI workflow does not start with “What should I post?” It starts with a real business constraint and ends with a measurable operating decision. We also learned that language models are strongest when they understand intent, plan actions, create content, and explain results, while dedicated numerical tools handle forecasting, lift, simulation, and optimization. Finally, trust features such as exact approvals, visible uncertainty, evidence quality, audit trails, and safe failure recovery do not have to make a product harder to use. When presented clearly, they make the AI feel more reliable and give owners greater confidence in acting on its recommendations.

What's next for Mercora

Next, we plan to turn Mercora into a production platform for real small businesses. That means adding persistent customer accounts, live uploads, real website publishing, direct marketing integrations, POS and booking connections, and outcome tracking across multiple locations. We also want to improve the intelligence layer with stronger forecasting models, business-specific learning, experiment support, and more accurate lift estimation from real customer data. The goal is for Mercora to learn from every campaign, owner decision, and business outcome. Our immediate focus is launching paid pilots with local bakeries, salons, studios, and retailers. Those pilots will help us validate the product, measure real impact, improve reliability, and build Mercora into a trusted AI growth and operations system for small businesses.

Built With

• codex
• css
• fastapi
• gpt-5.6
• html
• javascript
• neondb
• node.js
• openapi
• pydantic
• python
• uvicorn