ProcureMate
Multi-Agent AI · Procurement · Hackathon Entry · Working Prototype

ProcureMate

Two AI agents in one Switch room that find and check Nigerian suppliers for a buyer, who makes every decision.

2 AI agents Buyer decides Open source · MIT

Submitted to the Flip the Switch hackathon on Nov 6, 2026. Runs locally and includes a demo mode with fictional data, so it can be tried without an API key. Not deployed.

TimelineNov 1 to 6, 2026
CategoryMulti-Agent AI
TypeHackathon · Solo
PlatformSwitch + Mattermost
DeploymentLocal · Demo mode
Status● Working prototype

What ProcureMate does.

ProcureMate puts two AI agents in one Switch room to help a Nigerian buyer find and check suppliers. Scout searches and reads supplier sites. Vera re-checks the best candidates and says what she could not confirm. The buyer makes every decision, and the agents send nothing.

It is my own product idea and my solo entry to the Flip the Switch hackathon, hosted by WeMakeDevs and DevTools Academy with Switch AI as sponsor, in the Best Business Workflow track. There is no client.

Finding suppliers you can trust is slow.

Small businesses and office managers in Nigeria who need to buy goods, such as office chairs, spend hours searching for suppliers and still cannot tell who to trust. Many suppliers are hard to check online, and many sell only through Instagram, Jiji or WhatsApp.

ProcureMate does not pretend to solve trust. It gathers the evidence, sorts it by strength, and is explicit about what is unknown, so the buyer can decide.

A workflow with a human at the end.

The hackathon asked for a workflow where people and AI agents get work done together. A supplier check fits that well. Each agent has a clear job, and the judgement calls belong to a person: which suppliers to shortlist, whether a company register entry checks out, and whether a quote request goes out. That split is the product.

How a request moves through the room.

Two Claude Code agents, both running Claude Sonnet, share one Mattermost room run by Switch. They work in separate folders and cannot share files, so they hand work over with an @mention. The buyer stays in the same chat from request to quote.

Highlighted steps are the buyer's.

01
Buyer posts a request
A plain message in the room, for example "10 office chairs delivered to Lagos, ₦1.5 million, within 7 days".
02
Scout searches
Asks one question, writes the task down in the room, then searches with the Brave Search API.
03
Scout sorts and reads
Sorts results into real supplier websites and leads (directories, marketplaces, social pages, overseas shops), then reads each supplier's own pages for contacts, prices and delivery claims.
04
Handoff to Vera
Scout hands the best candidates to Vera with an @mention.
05
Vera re-checks
Compares name, address and phone across pages, and looks for a registration number on the site, an independent listing and the domain age. Groups suppliers by evidence and says what she could not confirm.
06
Buyer checks the company register
The buyer checks the Nigerian company register (CAC) by hand. Vera records the result as buyer-reported.
07
Buyer decides
Shortlists, excludes or picks suppliers in the same chat. Scout keeps a decisions list.
08
Buyer sends the quote request
Scout drafts the quote request and a ready-to-send Gmail link. The buyer sends it. The agents send nothing.

See it run.

A 2:40 demo of ProcureMate running in a Switch room. Open it on YouTube.

What it did in testing.

69
Tests pass in the public repo, with CI on Python 3.10 and 3.12
14 to 19
Nigerian supplier sites found in each real run
~6 min
From request to Vera's reviewed list in a Switch room, with some pages cached

The recorded run took about 11 minutes through to the quote links, and an earlier first run took about 18. A full live run used about 11 Brave queries, roughly 5 cents by the tool's own estimate. Few suppliers listed prices and none showed a confirmed Abuja address, and the agents said so instead of guessing. Four quote request emails were sent to suppliers during testing. The buyer sent each one through the Gmail link, not the agents.

Demo mode ran from a fresh download in PowerShell, Git Bash and WSL Ubuntu with no API key, and an audit found no secrets in the git history and no network calls in demo mode.

These are prototype results from a solo hackathon build. I make no claims about users, supplier replies, accuracy rates or time saved against manual search, because none of those were measured.

What was hard to build.

01
Honest evidence
Agents tend to overstate. Every fact carries a source, a time and an excerpt. Missing information is marked unknown, never treated as fraud. "Based in" and "delivers to" are separate claims, a supplier's claim stays a claim, and Vera never says "verified". A classifier keeps directories, marketplaces and overseas shops out of the supplier count.
02
Two agents in one room
A plain message reached no agent, so Scout is the main agent and follow-ups are turned off. The agents work in separate folders and cannot share files, so Vera runs her own checks from Scout's domain list and both use one shared page cache.
03
A flaky connection
DNS timeouts on a phone hotspot made agents report healthy sites as unreachable. A network error is now its own state, with retries, a slower final pass, a 24-hour cache and spaced visits. Long runs save progress and can resume.
04
Messy real sites
Numbers like "(+234) 0907…" and "0907-878-8888" are one number, so phone numbers are normalised and deduplicated. A student desk set called "Ergonomic…" was misread as an office chair, so product type is classified separately from price.
05
Safe and reproducible
A shared Brave query cap, a result cache, and a demo mode with fictional data let anyone run it without a key. Tests pass with the network blocked, and nothing is sent without a clear yes.

The stack behind it.

AI and agent platform
Claude Sonnet in Claude Code: Scout and Vera Switch Console and Switch rooms
Frontend
No custom UI Mattermost chat, started by Switch Console
Backend and data
Python 3, standard library only pytest Brave Search API RDAP for domain age JSON files and local cache, no database
Hosting and tooling
Local, Windows 11 with Docker Desktop Gmail compose link, no email provider Git and GitHub (public, MIT) GitHub Actions CI Clipchamp and ffmpeg for the demo video

Planned improvements.

A shared server so teammates can join the room. Email sending behind explicit per-message approval, with quote replies tracked. Manual lead lists for Instagram, Jiji and WhatsApp-only sellers. District coverage beyond Lagos, Abuja and Kaduna. Evidence shared in room documents, with agent roles set in Switch. Automated company register checks, only if CAC permits it.

What I bring to a team.

ProcureMate was built solo in about six days, and most of the work was in the rules around the agents rather than the agents themselves. Every fact carries its source, unknowns stay unknown, and a person makes each decision.

ProcureMate demonstrates designing agent workflows that are honest about their evidence and leave the final call with a human.