RASTO: from batch to dossier in 30 seconds

HACKATHON PROJECT PROPOSAL · MUROPLÁS

RASTO: from batch to dossier in 30 seconds

RASTO is a traceability assistant for Muroplás. In natural language and with validated data, it starts from a customer complaint, with the batch and the part sold to the customer, and answers a question the medical industry asks all the time: "show me everything that happened to this batch".

There are three reasons in favour of this topic:

• In injection moulding of medical and hospital components, traceability weighs heavily on the customer relationship.
• It is an easy story to tell a jury: a customer complains about a part from a batch, and the answer takes hours.
• It builds on what was already planned: an MCP server over the factory data, a validation layer and a demo agent.

What it does

1. Tracing from the complaint. Starting from the complaint, with the batch and the part sold to the customer, it finds the production orders, the machines and moulds, and the process parameters, inspections and non-conformities involved.
2. Complaint dossier. Generates a document with the batch's process parameters, inspections, non-conformities and shipping. It flags gaps, for example "no inspection record for the 2nd shift".
3. Validation before answering. Deterministic rules check for missing records, out-of-tolerance parameters and inconsistent dates. The local Laya can act as an optional semantic validator, with the rules as a fallback. Numbers and links always come from queries and rules.

How it scores on the 4 criteria

Criterion	Weight	What the project delivers	Proof in the demo
Innovation and originality	30%	"Agent-ready" factory data: the MCP server returns only validated facts and states its confidence level. It runs locally, so the data never leaves the factory.	The same server used by two different agents (yours and Cursor).
Technical implementation	30%	Traceability graph traversal done in code, typed MCP tools, read-only DB, validation rules, tests and an evaluation set with questions and expected answers.	An injected failure that the system catches, and a hit rate on the evaluation questions.
Impact and potential	20%	Fewer hours spent preparing dossiers and answering complaints, and faster, better-grounded replies to the customer. Applicable to any injection moulder.	Side-by-side timing: manual work in SSMS versus RASTO.
Presentation and demonstration	20%	A story with a beginning, middle and end, and a visual graph that needs no explanation.	5-minute demo and a backup video.

Architecture

MSSQL (read-only views) → MCP server in C# with tools such asrastrear_lote,parametros_do_loteandnao_conformidades→ validation layer → demo agent and graph page → dossier generator in Word or PDF.

2-day plan

Person A: data and server

• Day 1, morning: connection, views and anonymised data.
• Day 1, afternoon: MCP tools and validation rules.
• Day 2, morning: tests and evaluation set.

Person B: agent and presentation

• Day 1, morning: slides and pitch outline.
• Day 1, afternoon: demo agent.
• Day 1, end of day: graph.
• Day 2, morning: dossier.

Both

• Day 2, afternoon: code frozen 3 hours before the end, backup video and two timed rehearsals.

To build faster with AI

• Use Grok to generate the C# server skeleton, the tests and the validation rules from the DB schema and deploy.
• Use Grok to generate Presentations and simulations

The 5-minute demo

1. "A customer complained about part P from batch X. What happened to it?" The graph lights up and the path appears in seconds: production orders, machine, mould, parameters and inspections.
2. You request the dossier for a batch. The document comes out with a flagged possible problem.
3. You show the same server working in another agent.
4. You close with the time taken, compared with the manual process.

Risks and cuts

• The biggest risk is the DB not having all the links. In the first 2 hours, confirm that the links complaint (batch + part sold) → shipment → finished batch → production order → machine and mould exist. If they are missing, build a synthetic dataset with the same schema and say so in the pitch.
• Real data in front of the jury. Use anonymised data and confirm with the company what may be shown.
• Do not promise regulatory compliance. Present RASTO as a traceability aid and not as a certified system.
• If time runs short, cut the nicely formatted PDF, the second agent and the issue prioritiser. The core is tracing, dossier and validation.

Discarded alternatives

• Scrap or parameter-drift prediction: needs historical machine time series and model tuning, which is risky in 2 days.
• Mould-change assistant: a good idea, but with less data available and a smaller effect in the demo.