Skip to content
LyftTransportationNorth America

Lyft: Vision-Based Parking Compliance Automation

Replace repetitive visual review with production computer vision that stays controlled through real-world change.

Delivery roleData Scientist, model delivery

Context

Lyft micromobility operations in North America, where whether a vehicle had been parked correctly was decided from ride-end photos by human reviewers at operating scale.

Constraint

Lyft needed to determine whether vehicles were parked correctly from ride-end photos without relying on human reviewers at operating scale. The model had to handle real-world variation in camera angle, lighting, background, positioning, and occlusion, then continue to work when the physical scooter design changed during deployment.

Production decision flow
From ride-end photo to a trusted decision
Reliability through change

Automation handles clear cases, uncertain cases remain controlled, and the lifecycle accounts for model drift when the physical product changes.

System

Built a Keras CNN that classified ride-end photos and integrated it into Lyft's existing ML platform and deployment workflow. When the scooter hardware changed and the original model degraded, adapted the vision pipeline for the new design. Validated the automated decisions against the established human-review baseline through controlled A/B experiments, then introduced the system through progressive deployment rather than a single all-at-once release.

Outcome

Replaced a manual parking-compliance review workflow with scalable computer-vision automation. The model was adapted through a hardware generation change and was tested against real operational review decisions before wider production rollout.

How it was built

Approach

  1. 01
    Detection model
    Built a Keras CNN for image classification, trained to determine parking compliance from ride-end photos and integrated into Lyft's existing ML platform and deployment workflow.
  2. 02
    Model adaptation
    When the physical scooter design changed, the existing model, tuned to the previous hardware, degraded significantly; adapted the vision pipeline to recognize the new design while preserving performance across the broader range of parking scenarios.
  3. 03
    Benchmark validation
    Used historical human reviews as a reference set, then ran controlled A/B experiments comparing model-driven decisions against the established manual-review process.
  4. 04
    Statistical validation & rollout
    Applied statistical testing to confirm automated decisions held up against the human-review benchmark, then released progressively before broader production use.
Direct booking, no form required

Discuss a similar initiative

30 minutes to map the decision, the data behind it, and the shortest credible path to a production outcome.

Book a 30-minute callor email info@yannickbounouar.com
Back to work ·View services·Contact