Sr. Data Scientist, Programmatic Algorithms

Design and own the data infrastructure that feeds programmatic models: event pipelines, feature stores, training datasets, and real-time feature serving.

Engineer high-signal features from auction logs, bid stream data, user signals, contextual attributes, and historical performance — at the scale of programmatic data volumes.

Build robust data pipelines with production-grade standards: reliability, observability, versioning, and efficient reprocessing.

Deploy models to production real-time inference environments; own latency, reliability, and throughput requirements for auction-time decision-making.

Build monitoring systems that track model performance, data drift, and system health in production; define alerting thresholds and retraining triggers.

Partner with MLOps and Platform Engineering to ensure scalable, low-latency serving infrastructure meets SLOs under high-volume auction traffic.

Own the full model lifecycle: training, evaluation, deployment, A/B testing, and iteration.

Design and execute rigorous A/B and holdout experiments to measure the causal impact of model changes on yield, fill rate, advertiser performance, and publisher revenue.

Build evaluation frameworks that go beyond offline metrics — validating model behavior in live auction environments where feedback signals are delayed or noisy.

Translate experimental results into clear business narratives; present findings and recommendations to Product and business stakeholders.

Research and implement adaptive, self-learning components within the programmatic stack — including contextual bandits, reinforcement learning signals, and online learning approaches where appropriate.

Design feedback mechanisms that close the loop between auction outcomes, model updates, and system behavior; reduce reliance on manual tuning and rule-based overrides.

Stay current with advances in programmatic ML, auction theory, and online optimization; evaluate applicability to Impact's specific marketplace dynamics.

Serve as the primary ML technical partner for the Rubicon product and engineering teams; translate business requirements into modeling approaches and communicate technical tradeoffs clearly.

Collaborate with Data Science peers on shared infrastructure, modeling standards, and cross-domain feature reuse.

Document models, architectures, and experimental findings to a standard that enables review, replication, and knowledge transfer across teams.

Experience: 5+ years in data science, ML engineering, or quantitative research, with at least 2+ years building and deploying ML models in programmatic advertising, ad tech, marketplace optimization, or a closely related domain (e.g., real-time bidding, dynamic pricing, auction systems).

Programmatic & marketplace depth: Demonstrated understanding of programmatic auction mechanics (RTB, header bidding, floor pricing, deal types, bid shading) and how ML can be applied to optimize outcomes across the supply-demand stack.

Production ML engineering: Proven ability to take models from prototype to production independently — including real-time inference, monitoring, retraining pipelines, and SLO ownership.

Data architecture: Experience designing and building data pipelines, feature stores, and training infrastructure for high-volume, low-latency ML systems.

Technical skills:

Strong Python and SQL; proficiency with ML libraries (scikit-learn, XGBoost, LightGBM, PyTorch/TensorFlow) and large-scale data tools (Spark, Kafka, or equivalent streaming/batch frameworks).

Experience with real-time feature serving and low-latency model deployment (REST APIs, gRPC, or streaming inference).

Familiarity with production ML workflows: model versioning, drift monitoring, A/B testing, evaluation, and retraining.

Experience processing and modeling at programmatic data scale: high-cardinality auction logs, bid stream data, impression and click events.

Experimentation rigor: Strong grasp of causal inference and experiment design in online, delayed-feedback environments (auction holdouts, switchback tests, variance reduction techniques).

Communication: Ability to explain complex modeling decisions and tradeoffs to Product and business stakeholders; comfortable presenting in cross-functional forums.

Education: Bachelor's in a quantitative field (CS, Statistics, Math, Engineering, Economics, or similar); Master's/PhD preferred.

Direct experience with SSP, DSP, or exchange-side yield optimization — particularly floor price optimization, bid landscape modeling, or deal matching algorithms.

Familiarity with auction theory (first-price vs. second-price dynamics, optimal reserve pricing, revenue equivalence) and its practical implications for programmatic ML.

Experience with contextual bandits, multi-armed bandits, or reinforcement learning applied to real-time decisioning problems.

Knowledge of online learning and adaptive algorithms in production environments with non-stationary data distributions.

Familiarity with privacy-preserving ML techniques relevant to programmatic (differential privacy, federated learning, cookieless attribution modeling).

Experience with GCP tools (BigQuery, Vertex AI, Dataflow, Pub/Sub) and/or Databricks/Spark for large-scale event processing and model training.

Exposure to supply forecasting, inventory management, or capacity planning in programmatic or marketplace contexts.

Familiarity with Impact's affiliate and partnership ecosystem, or prior experience at the intersection of performance marketing and programmatic delivery.

Marketplace intuition. You understand programmatic auctions not just as an engineer but as an economist — you think about incentive structures, equilibrium dynamics, and how model decisions ripple through the supply-demand stack.

Full-stack ML ownership. You're as comfortable designing a feature store schema as you are tuning a gradient boosting model or debugging a latency spike in production. You own the whole chain.

Feedback loop thinking. You don't just deploy models — you design the systems that make them smarter over time. You think about how today's decisions become tomorrow's training signal.

Rigor under real-world constraints. You know how to run clean experiments in environments where feedback is delayed, data is noisy, and business pressures create tradeoffs. You don't let imperfect conditions become an excuse for imprecise thinking.

Pragmatic delivery. You ship. You balance the perfect with the production-ready, iterate fast, and know when an MVP outperforms a six-month research project.

Collaborative depth. You build genuine technical trust with engineering and product partners — not just by having good ideas, but by following through, communicating clearly, and making the integration easy.