Oscilla
GitHubSign in

AMD Developer Hackathon: ACT III · Intelligent Industry

Know which warehouse machine is wearing out, and what to do about it.

Oscilla reads the vibration of forklifts, AGVs and drones, scores each machine against its own healthy baseline, explains the likely cause from the manuals and the service journal, and turns the answer into a work order a technician can act on.

Try the demoSee how it works
Sample from the demo data
MachineStatusScore
Forklift FK-02Critical36
AGV Robot R2Needs attention53
Scan Drone D1Healthy86

The problem

Wear is found late

Warehouse equipment is serviced on a calendar or after it breaks. Early bearing wear does not show up in a walk-around.

Alerts do not say why

A single threshold on one sensor tells a technician that something is off, not what to inspect or how urgent it is.

Knowledge is scattered

What a machine did last month sits in a manual, a chat thread or a person's memory, not next to the reading.

How it works

  1. 01

    Read the signal

    A window of raw vibration becomes a handful of features: overall level, spikes, energy per frequency band and the envelope at the bearing defect frequency.

  2. 02

    Score against the machine's own baseline

    Each machine is compared with what it looks like when healthy. The result is a score from 0 to 100 and the features that moved, not just a number.

  3. 03

    Explain with sources

    The Maintenance Copilot gives the likely cause and next steps from the signal evidence, the manuals and the journal. Citations, numbers and durations are checked against the data. A critical machine always asks for human review.

  4. 04

    Turn it into work

    A recommendation becomes a work order with the AI diagnosis attached. Managers assign it; technicians take it, start it and close it.

  5. 05

    Learn from the technician

    When the job is closed the technician says whether the diagnosis was right. The overview shows how often the AI was confirmed.

One system, four jobs

  • Technician

    My tasks: what is assigned to me, what I can pick up, which machines to watch. Start, finish and write the journal.

  • Warehouse manager

    Overview of the warehouse, work orders with assignment, audit log and the state of every data source.

  • Administrator

    All warehouses side by side, people and roles, everything a manager sees.

  • Viewer

    Read-only overview and machine health, with no action buttons.

The same boundaries are enforced in the database with row-level security, so a screen that is hidden is also a door that is locked.

What is real and what is simulated

Real
  • Feature extraction, baseline scoring and the stored baselines
  • An ingest API a gateway can send raw windows to
  • Database, row-level security, assignment guard and audit trail
  • Copilot and agent guardrails: checked citations and numbers, human review, proposals only
Simulated in this demo
  • The vibration signals: dummy sensors send windows through the same scoring and storage path
  • The service manuals: short synthetic samples written for the demo
  • The data set: 16 machines in two fictional warehouses

Evidence

9 of 9fault recordings flagged
0 of 10held-out normal windows raised an alarm

CWRU bearing dataset, 12 kHz drive end, 0 hp, nothing tuned on this data. It is an easy test: one load and 10 windows per recording. Field data and remaining-life validation are still to do.

Built to run on AMD

The Copilot and the agent run on Google DeepMind's open Gemma 4 through an OpenAI-compatible endpoint served on AMD Token Factory, so the same code also runs with vLLM on an AMD Instinct MI300X (ROCm). Scoring runs as a small stateless service whose interface does not change when the model does.

Built with

Next.js 16 · React 19 · shadcn/ui · Tailwind CSS · Supabase (Postgres, RLS) · FastAPI · NumPy / SciPy · vLLM / ROCm

Team

Welcome World

Welcome World, for the AMD Developer Hackathon: ACT III, track Intelligent Industry.