SalesPal
A real-time e-commerce analytics platform utilizing the Vercel AI SDK and tool-calling pipelines to let merchants query store metrics using natural language.
Problem & Context
Traditional e-commerce dashboards (Shopify home, WooCommerce analytics) fail merchants when they need answers beyond pre-built charts. If an owner wants to know why conversion dropped on mobile last Tuesday among returning customers from specific campaigns, they are forced to export CSVs, write custom SQL, or click through a dozen disjointed reporting menus.
Store owners don't need more static dashboards; they need a data analyst they can talk to. SalesPal was built to replace rigid UI reporting with conversational intelligence — allowing merchants to query their live store performance using plain English and receive instant, grounded insights.
Product & How It Works
SalesPal is engineered around a secure, low-latency streaming architecture powered by the Vercel AI SDK.
[ Merchant Chat Prompt ] ──> ( Vercel AI SDK Stream )
│
▼
( Tool-Calling Engine )
│
┌───────────────────────┴───────────────────────┐
▼ ▼
[ Fetch Live Store Metrics ] [ Query DB & Aggregate ]
│ │
└───────────────────────┬───────────────────────┘
│
▼
( Strict Grounding & Guardrails )
│
▼
[ Grounded AI Response ]
Data Ingestion & Metrics Layer
Connects directly to e-commerce store databases to normalize orders, customer cohorts, traffic, and revenue streams into an accessible query layer.
Tool-Calling Architecture
The AI model does not calculate numbers from memory. Instead, it utilizes structured tool definitions to query live database endpoints, ensuring 100% mathematical accuracy.
Strict Guardrails & Grounding
System instructions prohibit estimation. If data is missing or out of range, the model falls back to explicit database queries or reports data unavailability rather than hallucinating metrics.
Key Features
-
Conversational Metrics Queries
Ask complex questions like "Compare our gross margin this week against last month's Friday peak" and get immediate answers with inline charts.
-
Automated Store Health Summaries
Proactive AI-generated digests highlighting anomalous drop-offs in conversion or spikes in cart abandonment.
-
Interactive Data Visualization
Dynamic UI components that render charts, trends, and tables directly inside the chat stream.
-
Lightning-Fast Streaming
Real-time token and component rendering leveraging Vercel's edge infrastructure.
Results & Metrics
Maintained an average response and data-fetch latency of under 1.2 seconds per conversational turn.
Reduced time-to-insight for store performance checks from 15 minutes of manual filtering down to a single conversational prompt.
Successfully proved the viability of using Vercel AI SDK tool-calling for deterministic financial and sales data reporting.
Tech Stack Deep Dive
- Next.js
- Full-stack framework powering both the reactive frontend and API route execution handlers.
- Vercel AI SDK
- Core library utilized for streaming text responses, managing chat state, and orchestrating server-side tool calls.
- TypeScript
- Ensures strict typing across data ingestion models, tool schemas, and UI components.
- Tailwind CSS
- Utility-first CSS framework for rapid, responsive UI design.
- shadcn/ui
- Accessible, beautifully styled component primitives used for dashboards, chat windows, and data tables.
My Role & Ownership
- Role: Technical Co-Founder & AI Architect
- Ownership: Personally architected the entire full-stack application, designed the database schema, built the Vercel AI SDK tool-calling pipeline, and integrated the reactive frontend components.
- Current Stage: Developed through active feature validation and iterative prototyping on preview environments.
Links & Resources
- Production Domain: N/A — currently running on preview infrastructure; custom domain setup pending public re-launch.
- Public Repository / Demo: Available upon request / private architecture review.
Lessons Learned
- 01
Never Let LLMs Calculate Math Directly
Exposing database queries via secure tool-calling is mandatory for analytics products. Allowing an LLM to "estimate" revenue figures destroys merchant trust instantly.
- 02
Streaming UX is King for Dashboards
When tool-calling takes 800ms to fetch database rows, streaming visual skeleton loaders and intermediate status messages keeps the UI feeling instant.
- 03
What I'd Do Differently
I would implement a proactive alert system earlier so the AI pushes critical metrics to merchants via email or webhook rather than waiting solely for user prompts.