Skip to content
All work

Maridian

Six Claude agents that recommend what to do with a flagged bakery defect, for line operators.

Role
Infrastructure lead and coordinator
Status
Built in 5 days
Event
TMLS Agentic Hackathon 2026
Team
5 people
Year
2026
Stack
Claude Haiku 4.5Anthropic SDKPythonFastAPISQLiteJavaScriptRailway

Problem

Cameras on a bakery line can flag defects at line speed, but what happens after the flag is still manual. A QA technician pulls the unit, classifies it, checks specs, phones a coordinator, works out the recovery value and picks an outcome.

My role

Infrastructure lead and coordinator on a team of 5. I set up the monorepo, the Vercel and Railway deploys, and the API contract, where a customer email draft can never be marked ready to send. I also redesigned the operator portal shown here; teammates built the agents, tools and database.

How it works

Six agents run in a fixed order, each a Claude tool-use loop over a seeded SQLite database, with no agent framework. They recommend; the operator decides.

1

Run the analysis

The operator opens a flagged line and runs the analysis; the API builds a defect event from that batch’s QC data.

2

Six agents reason

Triage, Options, Scorer and Decision pick an action; Batch Tracker checks the SLA; Recovery Planner runs only on a breach.

3

A person decides

Each agent’s steps stream to the screen; the operator accepts or overrides, and customer emails stay drafts.

The product

Maridian’s line operator portal on LINE-3: an Escalate recommendation marked pending, its written reasoning, and the agent reasoning panel with the Triage and Options steps.
One run on LINE-3: the recommendation, its reasoning, and each agent’s steps.
A partial-shipment recovery proposal above a customer fulfillment draft marked “Draft · pending account manager”, with the Send to customer button disabled.
On an SLA breach, the customer update is only a draft; a person has to send it.
The Production Lines overview: 2 of 7 lines running, 5 needing action, 12 active problems and 4 SLA breaches, with each line’s product, defect rate and status.
The overview: every line’s defect rate and SLA status.

Demo day

A handheld shot past people standing at the team’s table, coming to rest on its sign: “TMLS · Team #16 · Maridian · Scan to view project”, the QR code blurred, in front of two open laptops.
Mahyar, in a black T-shirt and a “MAHYAR 16” name tag, talking with a visitor in a crowded room.
Mahyar, seen from behind, talking with two teammates at their laptops.

Where it is now

We demoed it live at the TMLS Agentic Hackathon 2026 on May 29. The hosted API is gone now, so the demo no longer loads; the code is on GitHub.

  • 32.6 sOne full analysis, all six agents
  • 11Tools the agents call, over 14 tables
  • 245Commits in 5 days, team of 5

Next project

In progress

Graph-Native Retrieval for Clinical LLM Agents

Directed studies, York Data Mining Lab