Responsible for adoption of new technologies and patterns across the product commercialization org (~200 individuals).In charge of the architecture, technical roadmap, and engineering excellence bar for ~5 engineering teams (~40 managers and ICs) in collaboration with ML, Data Science, Product, and Engineering Management partners.Also utilized by executive leadership for hands-on development on our most ambiguous, difficult, or critical initiatives.Focused on building scalable and reliable distributed systems, service platforms, data pipelines, and algorithms for: cloud-based orchestration and movement of a large fleet of robots (Autonomous Vehicles) in urban environments, AV dispatching & supply optimization platform (ride hail supply & demand matching), AV route planning (real-time vehicle routes, traffic, ETAs, pick-up drop-off optimization), cloud-AV workflow state machine representation, communication, and task execution, geospatial data and multi-sided marketplace platform.

Building scalable software applications today is extremely expensive, requiring the combination of dozens of different tools. The engineering involves countless arcane tasks that are far removed from actual application logic. The narrow focus of each individual tool, and the lack of a cohesive model of building end-to-end applications, causes the engineering cost to be orders of magnitude greater than it could be.We are building a new kind of software tool that radically changes the economics of software development. Its foundation is a general-purpose programming language implementing a new programming paradigm. This technology pushes the boundaries of what's possible in compilers, databases, and distributed systems. It's not just for the initial construction of an application, but also encapsulates deployment, monitoring, and maintenance. It implements the first truly cohesive model for building software applications – a set of abstractions that can build any distributed application at any scale with greatly reduced engineering cost.We raised $5M from Initialized, Kindred Ventures, Rogue Capital, Background Capital, Max Levchin’s SciFi VC, and Naval Ravikant.

rideOS is Sequoia Capital's first investment in autonomous vehicles (total funding $34M). We built a platform of data and APIs (navigation, routing, maps, logistics, dispatching, network optimizations) which offers a straightforward way for auto manufacturers, autonomous vehicle companies, and transportation network companies to create and manage their own autonomous ride-hailing networks.As a founding engineer, I designed and built various parts of the stack: autonomous vehicle routing engine, dispatch optimization engine, geospatial databases, geospatial data engineering, and cloud infrastructure.- Led team of engineers responsible for vehicle dispatching algorithms for real-world autonomous vehicle ride-hailing fleets.- Built high performance distributed vehicle routing engine architecture: sharded sub-graphs, fast memory accesses using graph node geo-proximity sorting, with on-the-fly underlying graph topology updates.- Built distributed processing pipelines for maps datasets (e.g. OpenStreetMap) using Beam, processing the entire world graph (conversion, simplification, verification, etc) in ~1 hour.- Built geospatial database for multiple usage patterns such large datasets with versioning (worldwide HD/3D maps and traffic data), low-latency datasets (realtime position and telemetry of ride-hailing fleets), blob/media datasets (worldwide map vector tiles). Design principles: minimal cloud costs, consistency, versioning, fast geospatial selects, high-throughput writes.- Led cloud infrastructure and related tools / best practices: multi-environment GCP, gRPC, protobuf, Kubernetes, Docker, Helm.- Built all CI/CD pipelines for ~20 engineers: Bazel rules for Java/C++/Python, distributed builds on GCB, automatic and gated deploys (dev, staging, production) for dozens of micro-services all deployed every hour with Concourse.

I was the overall technical lead for Uber's Mapping, Routing & Logistics group, a group of ~60 engineers. We made Uber independent of Google Maps by launching a higher performance platform.I also designed and built Uber's vehicle route planning engine. It powers Uber's worldwide navigation, time estimates, price predictions, multi-passenger matching, and general logistics-related optimizations.Uber has maps and location at its core. We built a technology stack that enabled us to provide efficient and reliable transportation at worldwide scale.I was responsible for the design, architecture, and technical reviews of the Uber Mapping, Routing & Logistics services and data processing pipelines. That included: online vehicle routing, time / distance / price prediction engines, address and point of interest search engines, real time road traffic models, real time vehicle sensor data analysis, in-app navigation for drivers, etc.

I built Uber's vehicle and road speed prediction models. Those historical and real time data processing pipelines power Uber's vehicle routing engines and dispatching algorithms worldwide.I also managed the Mapping, Routing & Logistics Data Engineering team from 0 to 12 engineers and led it through the creation of our vendor-agnostic map data schemas, distributed vehicle sensor data processing pipelines, point of interest and address conflation algorithms. The team is responsible for building accurate map/geo/logistics datasets and algorithms that support teams across Uber such as vehicle routing, navigation, map search, map visualization, cartography, dispatch, and operations. We used: AWS, Kafka, Cascading, MapReduce, Samza, Storm, Thrift.

LiveRamp (acquired by Acxiom) is an AdTech and Data as a Service company. I led Data Engineering, a team processing 200TB/day of online marketing data on an in-house 300-node 5PB cluster. Our team was responsible for: data analysis, scaling distributed systems infrastructure, in-house NoSQL engine, fast Bloom-filter-based Cascading/MapReduce/Hadoop joins and multi-combiners, Map-side-only joins, statistical data inference, cross-device matching, consumer identity graph analysis (30B nodes), data cleaning and normalization, pipeline automation, distributed log processing (4B events/day), PII security and privacy best practices.I also authored HankDB (http://github.com/liveramp/Hank) a fast and ridiculously compact distributed key-value database. Specifically optimized for massive batch writes (up to millions of writes per second per node in the cluster), and for very large datasets. Read queries are guaranteed to execute fewer than 2 disk seeks at all times, and to perform 1 network call on average. Disk write operations are strictly sequential to achieve high throughput updates. Block compression of values, and a differentiating on-disk Flyweight pattern provide compactness. HankDB is fully distributed and fault tolerant. Data is replicated, consistently hashed and re-allocated automatically. It horizontally scales to petabyte-sized datasets with trillions of records. HankDB has been powering large production systems at Rapleaf, LiveRamp, and Acxiom with 100.0% uptime over the last 7 years.

I focused on processing and analyzing large amounts of demographic and online marketing data, building distributed systems, and scaling big data operations on a 200-node cluster. I was responsible for high-performance MapReduce/Cascading pipelines, data inference, statistical data validation and cleaning, the design and implementation of our social influencer scoring mechanism, and of our Solr/Lucene-based distributed search engine serving 20 billion documents.

We (a team of 5 engineers) built a high performance web-scale search engine from the ground up, with a very low dollar cost per indexed document. We wrote in-house C++ implementations of our distributed file system, MapReduce framework, indexing pipeline, ranking functions, and distributed search engine infrastructure.I was personally responsible for web graph analysis, search result clustering, language identification, document template detection, query autocompletion, query spell checking. I crafted and optimized ranking functions. I designed algorithms fighting web spam and web directories. I built a framework to deploy, run, and monitor distributed services and MapReduce pipelines on our cluster. I created tools to visually inspect, analyze and compare ranking functions. I was also responsible for implementing our web front-end.

I designed the training architecture and training algorithms of Tesseract, a fast and accurate Optical Character Recognition (OCR) engine which is open source.I leveraged internal Google datasets and improved accuracy on English text by 20%. I internationalized Tesseract which enabled use cases from Google Book Search, Google Street View, Gmail Anti Spam, and more. I made the training architecture generic, introduced n-gram-based language models, and improved the scoring engine. I made Tesseract efficient and accurate on 5 European languages in addition to English.