All case studies
Active DevelopmentComputer Vision · AI Systems2025 — present
TrafficPulse
Traffic analytics platform that turns raw video into structured scene understanding, not just detections.
Role · Lead engineer
Overview
TrafficPulse is a traffic analytics platform built around scene understanding. The objective is to move beyond per-frame object detection toward a system that models actors, lanes, signals, and traffic rules well enough to surface events a human operator would recognize as meaningful.
Problem
Standard vision stacks emit bounding boxes and class labels. Traffic operators need semantics — who is doing what, where, and whether it violates the rules of the road. That gap between detections and understanding is where most naïve pipelines fail.
Solution
A layered pipeline: perception (detection, segmentation, multi-object tracking), scene modeling (typed relations between actors, lanes, and signals), and a reasoning layer that maps observations onto traffic semantics. Each layer has a stable interface so components can be replaced independently.
Tech stack
- Python
- PyTorch
- Hugging Face Transformers
- Pydantic
- PyAV
- pytest
- mypy
- GitHub Actions
Engineering notesWhat's shaping this build.
Design goals, philosophy, planned architecture, and where the project stands today. No fabricated benchmarks — only what's actually driving decisions.
Engineering notes
What's shaping this build.
Design goals, philosophy, planned architecture, and where the project stands today. No fabricated benchmarks — only what's actually driving decisions.
Design goals
- Understand scenes, not just detect objects.
- Keep inference honest at streaming rates on realistic hardware.
- Make the evaluation harness score scene semantics, not per-frame accuracy.
Philosophy
The interesting engineering sits between the layers — tracking identity through occlusions, calibrating camera geometry to real-world coordinates, and defining what a correct answer even is.
Planned architecture
Three cooperating layers: perception, scene modeling, and reasoning — served behind a FastAPI service with a thin React dashboard for operator review.
Current stage
Perception and tracking backbone; first pass at the scene representation; a single-camera dashboard for live review.