Blue Vision Labs (acq'd by Lyft)

What Was Built

Blue Vision Labs built visual positioning and AR Cloud technology that let phones localize against high-accuracy city maps and anchor shared, persistent AR. Lyft acquired the company because the same computer-vision stack could accelerate Level 5 mapping and localization: detailed 3D maps of cities from camera-phone imagery, not only expensive lidar rigs.

My work sat in the operational layer that made that promise real: getting capture hardware into vehicles, collecting useful street data, and improving the pipeline between field collection and downstream mapping.

Blue Vision Labs shared AR demo with two phones seeing the same augmented content on a city street — Blue Vision started with city-scale visual positioning for shared AR. Lyft acquired the team because that mapping stack also mattered for autonomous-vehicle localization.

My Role

I helped test, improve, and scale the mapping operation across the U.S. and U.K. The role expanded from field testing into sensor redesign, vendor and part selection, mounting decisions, data-ingestion optimization, fleet deployment coordination, and privacy/compliance work around international camera data capture.

The work required constant tradeoffs across hardware reliability, field conditions, cost, heat, capture quality, latency, driver behavior, privacy constraints, and downstream mapping usefulness. A collection system that works in a lab is not the same thing as one that works across cities, vehicles, routes, weather, tunnels, and real operators.

Blue Vision Labs camera phone mounted inside a vehicle for street-level mapping capture — The core operating problem was making camera-based mapping reliable in moving vehicles, then scaling that capture process beyond controlled demos.

Result

The program scaled from 3 city centers to 2 countries and helped create up-to-date city-scale 3D maps for simulation, localization, trajectory analysis, and autonomous-vehicle development.

The broader Lyft Level 5 program later released what it described as the largest publicly released dataset of its kind: 55,000+ human-labeled 3D annotated frames, data from 7 cameras and up to 3 lidars, a drivable surface map, and an HD spatial semantic map. My contribution was not research-model work; it was the practical collection, ingestion, and field-operations layer that makes city-scale datasets and maps possible.

Why It Mattered

Autonomous vehicles need real-world data at scale, but real-world collection is messy: heat, tunnels, bad lighting, mounting variation, driver behavior, privacy constraints, and hardware failures all affect whether the data is useful.

The value was turning ordinary fleet activity into a lower-cost mapping network that could collect useful data across markets without relying only on specialized mapping vehicles.

What This Proves

I can operate where software, hardware, compliance, and field operations collide. The pattern is the same one I keep repeating: take messy real-world data, make the system observable, and scale it without losing trust in the output.