James Westover

Research Engineer — Applied Science, ML Systems, Bayesian Modeling, Probabilistic Inference, Geospatial & Sensor Data

Profile

Research engineer and applied scientist bridging academic physics and large-scale production ML systems. PhD-trained physicist (UCF, University of Rochester) with deep quantitative foundations in statistical mechanics, quantum chemistry (DFT), and computational methods — transitioned into industry building probabilistic and ML-driven systems at Uber, HERE Technologies, and Sonobi. Production experience spans real-time geospatial ML inference for fraud detection, sensor fusion pipelines for autonomous driving, FPGA-connected stream processing, probabilistic data structures at web scale, and Bayesian statistical modeling with PyMC5. Currently applying research engineering skills at Westover Labs / Neher Data Services, where I design ONNX Runtime inference services, knowledge graphs with pgvector semantic search, and Bayesian sports modeling systems using Polars and DuckDB. Peer-reviewed publications in surface physics. Acknowledged contributor to asteroid family dynamics research via large-scale computational methods.

Research Engineering Competencies:

Bayesian & Probabilistic Modeling: PyMC5 for hierarchical Bayesian sports models with uncertainty quantification; probabilistic group overlap estimation using HyperLogLog and MinHash at 250M+ impressions/day; GPS multipath correction via probabilistic shadow modeling (Uber ShadowMaps); anti-fraud probabilistic scoring ($30M+ annual savings)
ML Systems & Inference: ONNX Runtime production inference services (FastAPI, 10x speedup vs HuggingFace); PyTorch model deployment pipelines; on-device ML for iOS via ONNX; real-time ML scoring for location fraud detection at 2–10M QPS; AWS SageMaker prototyping
Scientific Computing & HPC: Beowulf cluster administration (ROCKS, MPICH2, OpenMPI) at University of Rochester; Monte Carlo simulations; molecular dynamics (LAMMPS); electronic structure calculations (VASP, DFT); FORTRAN (77/90/95); MPI-parallelized batch computation; microsecond-precision hardware timing for Bose-Einstein condensate laser cooling experiments
Geospatial & Sensor Fusion: Uber H3 and S2 cell indexing for global geospatial ML; real-time GPS multipath correction (ShadowMaps); IMU + camera + LiDAR sensor fusion for autonomous driving HD map generation; ROS-based pipeline design for edge capture and cloud ingestion; PostGIS for spatial analytics
Streaming Data Pipelines: Apache Storm (real-time stream processing, Python bolts) + Kafka for 250M+ GPS events/day at HERE; Amazon Kinesis; FPGA sensor stream integration (HT4 mapping rig); Kafka-based lambda architecture for vehicle fleet analytics; real-time localization pipeline for autonomous vehicles
Statistical Analysis at Scale: A/B testing infrastructure for ride-sharing platform; experiment design for route optimization algorithms; simulation frameworks scaling from 500 to 10,000 trip scenarios for off-route detection validation; Redshift-based campaign analytics (250M+ daily impressions)
Data Engineering: DuckDB + Polars for columnar analytics pipelines; Apache Spark, Hadoop, Hive, Druid, ClickHouse; Google Bigtable; pgvector HNSW indexes for semantic similarity search; ETL pipeline design with robust validation and lineage tracking
Physics Background: Peer-reviewed publications in surface science (DFT investigation of beta-SiC surface; organic/metal interface electronic structure). PhD coursework in quantum chemistry, solid-state physics, and computational condensed matter. Strong foundation in statistical mechanics, classical mechanics, electrodynamics, and numerical methods

Selected Experience

Westover Labs / Neher Data Services — Berlin, Germany

Founder (Westover Labs) & CTO (Neher Data Services) · 2024 – Present

Design and build Bayesian sports modeling system using PyMC5 for hierarchical probabilistic models with uncertainty quantification — enabling principled forecasting under uncertainty with posterior predictive distributions. Stack: Python, PyMC5, Polars, DuckDB for high-performance columnar analytics pipelines.
Architect and deploy production ONNX Runtime inference services via FastAPI — achieving 10x inference speedup over HuggingFace baselines. Serve object detection and NSFW classification models with sub-100ms latency at production scale.
Build knowledge graph service backed by PostgreSQL + pgvector with HNSW indexes for semantic similarity search — enabling fuzzy entity retrieval across a growing corpus of technical knowledge. Full migration management via Alembic; dual REST + MCP API surfaces.
Design probabilistic DFS (Daily Fantasy Sports) optimizer using constraint programming and stochastic modeling for lineup generation under salary constraints and projection uncertainty.
Operate multi-model ML inference pipeline supporting free/paid tiers with dynamic model selection, batch processing optimization, and rate-limit-aware scheduling for iOS clients.
German-registered company (Westover Labs UG) operating fully remotely with EU and international clients.

Uber — San Francisco, CA / Remote

Staff Software Engineer · February 2021 – October 2022

Senior contributor on Uber's navigation platform — one of the company's largest distributed systems (~10,000 nodes per data center), powering all routing and real-time geospatial inference globally. Code still runs on 30M+ driver devices.
Developed Python-based simulation framework for off-route detection algorithm validation — scaled test scenario capacity from 500 to 10,000 trips per run using multiprocessing parallelization. Enabled data-driven tuning of navigation heuristics before production rollout.
Reduced p95 off-route detection time from ~5 minutes to ~2 minutes by introducing cross-street link support to the route object and enhancing navigation engine heuristics — a geospatial ML inference improvement with direct rider safety implications.
Diagnosed and coordinated cross-team remediation of critical GPS clock offset bug affecting data integrity across millions of devices globally — root cause analysis spanning firmware, OS, and backend ingestion layers.

Senior Software Engineer — Python Platform (Office of CTO) · January 2019 – June 2020

Became de facto Python platform support for Uber's data science organization — maintained notebook infrastructure, supported distributed query systems (Presto, Hive), and debugged Uber's internal Airflow fork for ML workflow orchestration.
Contributed to regulatory compliance infrastructure for Uber's London market expansion — platform tooling supporting market-specific ML feature flagging required by UK transport regulators.
Led enterprise-wide Python 2→3 migration spanning hundreds of microservices and 90%+ of production systems — built dependency graph tooling to identify migration order and automate compatibility patching across the ML and data science stack.

Software Engineering Manager — Location Platform · February 2018 – January 2019

Built Uber's Location Platform from the ground up — designed GDPR-compliant real-time GPS + sensor data ingestion pipeline, driving global data completeness from ~80% to 97%+. Improved data quality directly enabled the RideCheck safety ML system.
Launched the ShadowMaps project — probabilistic geospatial correction for GPS multipath error in urban environments using building shadow modeling and satellite geometry. Blog: Rethinking GPS.
Led anti-location-fraud program: designed probabilistic scoring system to detect GPS spoofing in real time, delivering $30M+ annual savings. Coordinated ML model integration with real-time Cassandra-backed scoring infrastructure at 2–10M QPS.
Scaled globally distributed Cassandra infrastructure — managed consistency model tradeoffs (QUORUM, LOCAL_QUORUM, ONE) for high-availability location data ingestion and fraud signal storage.
Built rider-driver location sharing system with H3 and S2 geohashing — engineered real-time geospatial indexing for privacy-compliant, low-latency location queries.

HERE Technologies — Berlin, Germany

Software Engineering Manager → Senior Software Architect · April 2015 – February 2018

Architected and built the Fleetvision Data (FVD) platform — a lambda architecture for scalable autonomous driving data ingestion and analytics using Apache Storm (real-time, Python bolts) + Kafka (ingestion) + Elasticsearch / PostGIS (spatial search and storage). Processed 250M+ GPS events/day from global mapping fleet.
Ran the largest single unsharded PostgreSQL/PostGIS instance in production — the entire HERE world map in one database, pre-HDFS-GIS era. Research-grade problem: spatial queries at APAC/EMEA/LATAM/NORAM partition boundaries for autonomous vehicles crossing shard edges in real time. Invented the approach because tooling hadn’t caught up yet.
Used Apache Thrift as universal IDL: one schema generated Python services, C++ ROS nodes, and JS types. Storm topology with Thrift-classified tuples — a typed streaming contract end to end across heterogeneous consumers.
Designed "Weatherman" — an orchestration system for dynamic map reprocessing and QA workflows. Kafka-based event-driven batch processing triggered downstream analytical pipelines for HD map generation and data quality scoring.
Integrated FPGA-based sensor stream encoding (HT4 next-generation mapping rig) — designed software abstraction layer for JPEG-compressed multi-camera streams from FPGA encoder boards, enabling reliable high-bandwidth ingestion into cloud analytics pipelines.
Built ADAS prototyping pipelines integrating Python processing with ROS (Robot Operating System) — rapid prototyping of perception and localization algorithms using real-time sensor data from mapping vehicles (GPS, IMU, camera, LiDAR).
Led Fleetvision 2: real-time fleet telemetry visualization platform with PostgreSQL/PostGIS, Elasticsearch, Redis, OpenLayers, and GeoServer — providing actionable spatial analytics for autonomous driving data collection operations.
Managed technology partnerships with Intel, Nvidia, Audi, and Mercedes-Benz for sensor-to-cloud ADAS and autonomous driving data infrastructure.
Built ISO 9001-certified quality management system for safety-critical mapping vehicle software — authored QMS documentation and led auditor engagement for software governing vehicles on public roads.
Scaled engineering team from 2 to 15 across backend, embedded, mobile, DevOps, and frontend subteams delivering 8+ concurrent ADAS and autonomous driving programs.

Sonobi — New York, NY

Software Engineer · December 2013 – January 2015

Designed probabilistic group overlap estimation system using HyperLogLog (cardinality estimation) and MinHash (Jaccard similarity approximation) — efficiently computed intersection sizes between large user cohorts without storing raw identifiers. Backed by Redis for HLL/MinHash object persistence; validated with statistical confidence testing via NumPy.
Built Warehouser — impression-level data warehousing ETL pipeline aggregating 250M+ daily ad impressions across campaigns using S3, Redshift, RabbitMQ, and Python. Enabled granular reporting and statistical optimization for campaign performance analysis.
Explored real-time log analytics pipeline with Kafka, Logstash, and Hadoop — early implementation of streaming analytics for ad server event data.
Built web scraping and API reverse engineering framework using Wireshark-based HTTPS traffic analysis for competitive bid pricing data collection.

University of Rochester — Rochester, NY

Assistant Systems Administrator · October 2006 – October 2012

Built and administered Beowulf-style HPC cluster using ROCKS with diskless node deployment — supporting MPI-parallelized scientific computing jobs (MPICH2, OpenMPI) for physics research groups.
Implemented and maintained private cloud infrastructure using Eucalyptus (open-source EC2) later migrated to VMware — providing research groups with on-demand compute for Monte Carlo simulations and electronic structure calculations.
Supported physics research computing workloads including molecular dynamics (LAMMPS), electronic structure calculations (VASP / DFT), and high-throughput parameter sweeps across cluster nodes.
Designed microsecond-precision real-time timing systems for laser cooling experiments in Bose–Einstein condensate research — hardware interfacing, signal synchronization, and LabView-based experimental control systems.

University of Central Florida — Orlando, FL

Graduate Teaching Associate & Researcher · August 2009 – December 2013

Conducted original computational physics research leading to two peer-reviewed publications in surface science: DFT investigation of beta-SiC(100) surface metallization, and electronic structure of pentacene/Cu(110) organic-metal interface.
Applied density functional theory (DFT) and electronic structure methods (VASP) to model quantum mechanical phenomena in condensed matter systems — integrating computational results with experimental observations.
Designed and taught lab curriculum for studio-format introductory physics using high-speed video analysis and student-driven experimental design — emphasis on measurement uncertainty, data analysis, and hypothesis testing.

Valencia Community College — Orlando, FL

Adjunct Faculty — Physics & Astronomy · December 2010 – May 2014

Taught introductory Physics and Astronomy; developed and implemented hybrid-mode physics courses integrating online simulation tools with in-person lab instruction.

Education

MBA, Business Administration — Quantic School of Business and Technology, 2017

PhD coursework, Physics (Quantum Chemistry, Solid-State Physics, Computational Condensed Matter) — University of Central Florida, 2009–2013