Mantis Analytics

Tierscope tells a manufacturer which suppliers could stop its factories, how much money is at risk, and what to do when one fails. Its planning agent runs on Meta's open-weights Muse Glimmer on the company's own GPU, so the supplier graph never has to leave the company's infrastructure. A car is built from thousands of parts, and every supplier has suppliers of its own. When one small supplier fails, every product that uses its part can stop, and nobody sees this from a spreadsheet. Tierscope reads supplier, part and factory files as they come (Excel, CSV, PDF, text or zip archives): a model maps each new layout once, and plain code reads every row. From them it builds a tiered map of who supplies what to whom and computes Business Exposure, the share of output that depends on one supplier, site or part, and Business Impact, the same risk in euros. Every value keeps its source cell or document quote, its confidence and its history, and every change can be undone. Facts the model is unsure of wait for a person, with a preview of how the risk numbers would change. When something goes wrong (a supplier fire, a closed shipping lane, a tariff, a chip shortage), entered with a form, picked from a library, pasted from a news alert or read from a document, Tierscope simulates it week by week and shows what it costs if nobody acts. The agent then investigates the graph, tests plans in the same simulator and proposes one. A person accepts or rejects every action; rejecting one with a comment produces a new plan version that follows it. Accepted actions become tasks, and each disruption ends in a report of what it would have cost, what the plan cost and what the model cost. We tested the final agent on Glimmer once, on cases it was never developed on (0 is doing nothing, 1 is an optimiser that knew everything in advance). On 22 single disruptions it scored 0.98, against 0.84 for a rule-based playbook. Over 100 simulated years of disruptions, seeing only each week's news, it scored 0.96 at a total cost of EUR 279M, against EUR 824M for the playbook. Built during the hackathon (Oct 27 – Nov 4): Tierscope (ingest, review, graph, exposure, web app); operations (disruptions, weekly simulator, response agent, decisions, tasks, reports); both benchmarks and the ShockBench exporter scripts. Existing before: ShockBench-Flow 0.1.2, Mantis Analytics' semiconductor-network benchmark, used unchanged as the event source; and a customer's data-capture template with its Business Exposure and Business Impact formulas, which Tierscope reproduces (not in the repository).