Vizwin — Retail Decision Intelligence

An analyst-led, AI-agent-backed platform that turns a small retailer's messy weekly business data into a human-verified, AI-optimized decision brief — in 1–2 business days.

Brand principle: Human First. AI Next. This is not "paste a spreadsheet into ChatGPT and trust the answer." AI drafts evidence-based claims; a human verifies every claim; the final report is built only from approved claims.

📌 This repository is a public case study. The production codebase is private (it holds client data connections and proprietary logic). This document covers the problem, the product, and the architecture.

---

The problem

Small and midsize retail operators (≈1–100 locations) sit on sales, inventory, promotion, and product data — but rarely have an analytics team to turn it into decisions. The two existing options both fail them:

• Hire an analyst / agency → too expensive and slow for a 3-store operator.
• "Upload a file, let AI decide" → confidently wrong, no accountability, no one will stake a decision on it.

Vizwin is the middle path: an affordable analytics department that pairs AI agents with human review at the start, where quality and trust matter most.

The product

A retailer uploads their weekly files. Behind the scenes:

1. Files are parsed and ingested into structured datasets.
2. An agentic Analyst Org computes findings with real SQL over that data — it doesn't guess.
3. Every finding becomes a claim carrying evidence, source, confidence, and risk.
4. A human reviewer approves, edits, or rejects each claim.
5. The report is assembled only from approved/edited claims and delivered as a decision brief with tracked action items.

Why it's different — the claim-verification model

The core insight is an enforced data invariant, not a marketing line:

• AI generates claims first, never the report directly.
• A claim with no evidence is automatically UNSUPPORTED and can never enter a report.
• REJECTED claims never enter a report. Only APPROVED / EDITED claims do.
• The report builder reads claims, not raw files — so every sentence delivered to a client traces back to verified evidence.

This makes the output defensible: an operator can act on it because a human already stood behind every claim.

Architecture

flowchart TD
    A[Retailer uploads weekly files] --> B[Parse + structured ingestion]
    B --> C[(Datasets — read-only SQL layer)]
    C --> D[Agentic Analyst Org<br/>domain specialists compute findings via SQL]
    D --> E[Tiered Claims<br/>evidence · confidence · risk · assumptions]
    E --> F{Human review}
    F -- approve / edit --> G[Approved claims]
    F -- reject --> X[Excluded — never in report]
    G --> H[Report builder<br/>approved claims only]
    H --> I[Decision brief + tracked action items]

Loading

Three surfaces, one workflow: the system proves the full loop from the client (signup, guided intake, multi-file upload), the admin reviewer (claim approve/reject/edit, report send), and report delivery (expandable evidence, report chat grounded only in approved claims, PDF, action tracking).

Engineering highlights

• Agentic analysis that computes, not hallucinates — specialist agents write read-only SQL over the structured data (model-written SQL is treated as untrusted), self-critique, and emit tiered claims with quantified prescriptions.
• Graceful degradation — with no AI key, the system falls back to a deterministic mock engine so the entire workflow still runs and demos offline.
• Evidence-bound report chat — answers come only from the report + approved claims; if the data isn't there, it says what's missing instead of inventing an answer.
• Security by default — server-only AI calls, signed httpOnly sessions, rate limiting on auth/upload, and a hard rule that no untrusted input reaches the database directly.

Tech stack

Next.js 14 (App Router) · React 18 · TypeScript · Tailwind · Prisma + Postgres (Supabase) · Zod · OpenAI (agentic, server-only) · Vitest

---

Status

Working prototype with seeded demo data and a live pilot deployment path. The model — AI drafts, humans verify, operators act — is enforced in the data model and covered by tests.

_{Built by Austin Stevens — Applied AI & Data Analytics, University of Tennessee, Knoxville. Source code private; this is a product/architecture case study.}