Agentic sourcing and outreach for an e-commerce wholesaler
A system that finds products worth stocking on Amazon, works out who supplies them and prepares the outreach. An agent runs it on a schedule, and a person approves every stage before anything moves on.
- Client
- E-commerce wholesale business name, if they agree
- Sector
- E-commerce, Amazon wholesale
- My role
- e.g. Solution architect and lead developer
- Built with
- Python, Airtable, agent orchestration on a cloud server
- Status
- e.g. In production since August 2026
products screened a week
of manual research saved a week
of outbound messages approved by a person before sending
The problem
Finding wholesale products that sell profitably on Amazon is research-heavy work. Someone has to shortlist products, check the numbers, work out which brands and distributors supply them, find the right contact and write to each one. Done by hand, that takes hours per product, and the quality depends on who did it.
Add the client’s own before-picture: how many products they reviewed a week and how long each took.
What I built
A sourcing and outreach system in three layers:
- A deterministic Python pipeline does the heavy lifting: collecting product data, scoring opportunities, matching suppliers and drafting messages. The same input always gives the same output, so results can be checked.
- An agent orchestrates the pipeline on a cloud server. It schedules runs, moves work between stages and reports on what happened.
- Airtable is the team’s control room: a live view of every product and supplier, and the approval desk where people review each stage before the next one starts.
- AI agent
- People
- Systems
Guardrails
- Human approval gates between stages, held in Airtable where the team already works.
- Deterministic code for scoring and data handling, so results are repeatable and auditable. The AI handles orchestration and drafting.
- Credentials injected through environment variables, never stored in code.
- Built and tested locally first, then deployed, with each module verified before the next one started.
How it was built
The build followed a dependency graph: each module was specified, built and validated before anything that depended on it. I used an AI-assisted engineering workflow with separate architect, builder and validator agents, and reviewed every change myself before it moved on. It’s the same principle the finished system uses.