I’m a machine learning engineer with 7+ years of experience. Currently, I’m on the foundation models team at Zoox, training vision-language-action models for autonomous driving.

I’ve previously worked on reinforcement-learning-based airline optimization at FLYR and micromobility data science at Lime and Lyft.

Because of the complex and interconnected nature of the systems I’ve worked on, I’ve learned to think flexibly about ML and leaned on approaches from supervised learning, reinforcement learning, representation learning, retrieval, simulation, time-series forecasting, causal inference, operations research, and user research.

I ask a lot of questions and am particularly good at designing evals that respect domain-specific concepts and concerns.

Github - LinkedIn - Instagram - Email

Language-aligned scenario representations are useful for driving scenario retrieval: https://www.nvidia.com/en-us/on-demand/session/gtc25-s73216/

Extrapolation error is the biggest challenge of off-policy RL. At FLYR we experimented with multiple ways of handling it: https://flyr.com/resource-hub/unlocking-efficiency-and-revenue-deep-q-learning-with-batch-constraints/

Specialized loss function for a neural network trained to output beta-binomial distribution parameters. Used for airline passenger no-show prediction at FLYR.

[Exploring the use of pseudo-labeling for commute trip detection at Lime: https://medium.com/lime-eng/identifying-commute-trips-at-lime-f10c79d7744f](https://prod-files-secure.s3.us-west-2.amazonaws.com/67709824-7583-4062-9f5e-36fb16832347/b4c120b7-70e1-4323-85cd-3d215fa96778/output.mp4)

Exploring the use of pseudo-labeling for commute trip detection at Lime: https://medium.com/lime-eng/identifying-commute-trips-at-lime-f10c79d7744f

Detecting helmets using computer vision models (1st place at Lime hackathon!).

Detecting helmets using computer vision models (1st place at Lime hackathon!).

Time-series forecasts for famine events, a project I consulted on at Milliman: https://openknowledge.worldbank.org/entities/publication/dd76937b-ff5c-5144-9df1-ea7fc1bf36fc

You can’t study reinforcement learning without having a favorite Atari game.

Lechner et al. were able to get really intriguing results for control-problems using very small biologically inspired neural network architectures. I tried training them in the Atari environment to get a better intuition.

The batter-pitcher interaction in baseball is a two-player, zero-sum game so it must have a Nash equilibrium according to the Minimax theorem. We examined the implications for optimal pitch-type selection and expectation.

Professional Experience

• Multimodal Foundation Models

  • Post-trained VLAs with SFT/RL to generate driving trajectories. Worked on our action expert architecture, training flow-matching fields with policy-gradient methods, data curation/pipelines, and developing evals with an emphasis on long-tail scenarios and non-Markovian behaviors. (TorchTitan, Ray, K8s, EKS, Slurm)

• Representation Learning for Driving Scenarios

  • Developed encoder architectures for learning language-aligned driving scenario representations that are useful for semantic retrieval. (Pytorch, Opensearch, Spark, AWS)

  Harnessing Generative AI to Scale the Autonomous Vehicle Industry S73216 | GTC 2025 | NVIDIA On-Demand

  Some applications of this work was summarized in this Nvidia GTC conference talk (16:22-18:49)

• Reinforcement Learning via Supervised Learning for Airline Pricing

  • Investigated the efficacy of replacing off-policy RL methods with conditioned imitation learning methods (similar to the decision transformer) in the airline pricing problem domain.

• Batch-Constrained Deep Q-learning for Airline Pricing

  • Helped develop an off-policy RL model that generates optimal flight pricing policies by learning action values from historical pricing and revenue data.

  Unlocking Efficiency and Revenue: Deep Q-learning with Batch Constraints

• Optimal Airline Overbooking

  • Developed and implemented in production a semi-parametric, deep learning-based approach to the airline overbooking problem. The two-step solution trains a neural network model to output distributions of passenger no-show probabilities and then uses a newsvendor style optimization approach to choose an overbooking level given the distribution and cost inputs. (Tensorflow, Beam, Airflow, Bigquery, Kubeflow, Streamlit, GCP)

• Self-Supervised Contrastive Pre-training for Airline Time-Series Prediction

  • Conducted research on the use of contrastive self-supervised objectives to learn rich feature representations of historical flight pricing dynamics. The representations were used in a range of downstream tasks to improve both performance as well as invariance across a few important dimensions.

• Simulating the Airline Revenue Environment

  • Contributed to the development of an internal simulation platform that enables the training and testing of reinforcement learning algorithms in the airline revenue management problem environment. (Jax, Evojax)

• Spatiotemporal Forecasting for Scooter Fleet Optimization

  • Developed component models for Lime’s scooter fleet optimization system. This included deep-learning models for downstream revenue prediction by geohash, statistical time-series models for forecasting repair task volume, and tree-based models for battery range estimation. (Keras, XGBoost, Scipy, Airflow, Snowflake, Tableau, AWS)

• User Segmentation Redesign

  • Developed an ML model for detecting commute trips.

  Identifying Commute Trips at Lime

  • Created the user segmentation framework used across the company by combining RFM and JTBD principles, pioneering commuter detection, and greatly improving tourist detection.

• Simulating the Micromobility Revenue Environment

  • Designed a simulation of the micromobility revenue environment and used it to identify candidate algorithm improvements for real world A/B testing.

• Causal Impact of adding Bikes to a Previously Scooter-only Market

  • Applied causal inference methodologies including synthetic control and coarsened exact matching to estimate the causal effect of offering multiple types of vehicles on rider behavior.

• Swappable Battery Tracking Pipeline and Metrics

  • Built out the central data pipeline and metrics used to track swappable battery assets.

• Computer Vision for Rider Helmet Detection

  • Worked on a helmet detection model (using YOLO-tiny-v4) for compliance in markets requiring riders to wear helmets. Won 1st place at internal hackathon.

• Stochastic Modeling of Food Insecurity

  • Implemented and assisted in the theoretical development of a stochastic time-series model for forecasting famine events at different spatial and temporal resolutions. (sklearn, scipy, R)

  Stochastic Modeling of Food Insecurity

• LLM Text-Embeddings for Improved Insurance Claim Cost Prediction

  • Improved an insurance claim cost prediction model by generating embeddings for text descriptions of accidents using BERT.

• UX Research on Pricing
  • Conducted user research to identify ways that average order value could be increased by changing the display of pass options during the mobile sign-up flow in Lyft/Motivate’s bike share app.
• Bayesian A/B Testing
  • Implemented a Bayesian A/B test to evaluate and communicate results of proposed product offerings change.

Projects

• Reinforcement Learning using Neural Circuit Policies for Atari Games

  • Experiments replacing the function approximator in PPO with liquid neural network layers that model differential equations instead of direct quantities and which use biologically inspired sparsity structures. (pytorch, gymnasium, ncps)

  https://github.com/nickdulchin/cleanrl

• Self-Supervised Learning for Foursquare POI Recommendations

  • Project from network economics class in college. Used the DeepWalk algorithm for recommendation.

  https://github.com/nickdulchin/foursquare-DeepWalk

• Baseball Research

  • I played baseball growing up and in college and it is what got me into statistics! Here are some projects that I worked on with the baseball analytics club at Tufts and when I was lucky enough to intern with the New York Mets.
    • Modeling the Pitcher-Batter Interaction with Game Theory (Saber Seminar, 2016)
    • Forecasting Player WAR with KNN (Diamond Dollars Case Competition, 2015)
    • Estimating the Value of MLB Draft Spots (New York Mets, 2014)
    • Examining the Potential Profitability of Team Specific Ballpark Alteration in Major League Baseball (MIT Sloan Sports Analytics Conference, 2014)