Skip to main content

Open Source Drone Recognition Pipeline

Classification Conference demo / ML systems pipeline

Status Built for talks

Period 2017 era

Materials Apache Flink, Apache Solr, Apache Mahout, eigenfaces, vector lookup, hacked toy drones

Problem​

Apache Mahout needed a concrete story for audiences who knew Apache Flink or Apache Solr but did not necessarily know why Mahout mattered.

So the demo made the system physical.

Build​

A set of inexpensive toy drones were hacked into mobile cameras. Their video fed a realtime pipeline using Flink, Mahout, and Solr. Mahout generated eigenface vectors. Solr handled vector lookup. Flink moved the data through the system.

  1. Inputtoy drone camera
  2. StreamFlink
  3. RepresentationMahout eigenface vector
  4. LookupSolr vector search
  5. Outputdemo result
demo goal: make the system legiblenot production recognitionthe wrapper mattered

What worked​

The talks were accepted, the demo drew attention, and the architecture became easier to discuss. The audience had a reason to engage with Mahout because the stack was visible, testable, and grounded in a concrete flow.

What broke​

The point was not to build a production-grade face recognition system. Eigenfaces had all the usual weaknesses: lighting, shadows, pose, rotation, and brittle matching. The point was to make the architecture tangible enough for a conference audience.

Lesson​

A technical demo has a job. Sometimes that job is not production accuracy. Sometimes it is making a system visible enough for people to care.

Why it matters now​

Agent work has the same pattern. The model is only one part. The interface, workflow, audience, trust boundary, and demonstration all shape whether the system matters.

Useful systems need more than a model. They need an interface, workflow, audience, trust boundary, and reason to exist.

Evidence​