The problem your project solves

Every ops and procurement team bleeds hours on vendor phone calls — chasing delivery updates, placing restock orders, negotiating payment timelines. These calls are repetitive, undocumented, and entirely manual.

The person making the call has to know the right questions to ask, remember to follow up, and then manually log the outcome somewhere. When stock runs low at 11pm, nobody is calling the supplier. When the call happens, the outcome lives in someone's head or a sticky note — not in a system.

The result: delayed restocks, missed follow-ups, zero audit trail, and ops staff stuck doing work that shouldn't require a human.

How you are solving it

Vendor Wrangler is an AI-native ops concierge that handles vendor calls end-to-end — the user types a request in plain English, and the system does the rest.

The pipeline:

Plan — Claude converts the natural language request into a structured JobSpec: objective, required questions, escalation guardrails, and an echo-mitigation prompt
Call — Genspark dials the vendor via PSTN and conducts the call using the JobSpec as its brief
Extract — Claude cleans the raw transcript and extracts a structured CallOutcomeSummary: resolution status, payment date, confidence score, and recommended next step
Save — The full report is stored in Supabase, scoped per user via RLS, and visible on the dashboard instantly
For suppliers, incoming customer restock requests appear as tickets with order status tracking (preparing → in_transit → delivered). Marking a delivery in-transit auto-checks inventory; marking delivered auto-increments the customer's stock levels.

For customers, a color-coded inventory dashboard (critical → warning → caution → healthy → abundant) surfaces low-stock items. One click triggers an AI restock call to the linked supplier — no form filling, no manual request.

Use of Genspark

Genspark's "Call for Me" API is the core execution layer — it places real PSTN calls to vendor phone numbers and returns a raw transcript.

The JobSpec generated by Claude is passed directly to Genspark as the call brief, including the objective, required questions, and escalation guardrails
An echo mitigation prompt is injected into every Genspark payload to prevent the AI caller from being manipulated by the vendor into going off-script
A USE_MOCK_CALL environment flag bypasses Genspark for safe demo environments while keeping the full pipeline intact
The raw transcript returned by Genspark is then handed to Claude for cleaning and structured extraction

Use of Claude

Claude (claude-sonnet-4-6) powers two critical stages of the pipeline:

1. Planning — Natural language → JobSpec
   The user's plain-English request (e.g. "Call FreshMart and ask when our A4 paper order is arriving") is sent to Claude, which generates a structured JobSpec containing: the vendor objective, a list of required questions, escalation guardrails, and an echo mitigation prompt. This ensures every call is purposeful and recoverable.

2. Extraction — Raw transcript → Structured outcome
   After Genspark returns the call transcript, Claude cleans it (removing filler, normalising speaker turns) and extracts a CallOutcomeSummary:

resolutionStatus — what was agreed
paymentDate — if relevant
confidenceScore — how certain the outcome is
nextStep — one of Needs Human Approval, Mark as Resolved, or Draft Email
restockItems — items and quantities confirmed for restock orders
Claude handles the full semantic reasoning so the dashboard always shows clean, actionable data rather than a wall of transcript text.

What is the deployed URL for this project?

https://vendor-wrangler.vercel.app/auth/login