Adaptive Gaming Experimentation
Human-in-the-loop analyst workbench for Dell: experiment validation, causal estimates, and next-best recommendations
Most teams can run experiments; fewer can reliably decide what to run next. The Adaptive Experimentation Agent closes that loop: it reads past test results,...
Problem: High-dimensional product and gaming experiments produce lots of data but little decision clarity. Analysts need a repeatable way to trust results and choose the next test, not another one-off dashboard.
Solution: The Adaptive Experimentation Agent is a human-in-the-loop analyst assistant: load experiment history → automated validation → causal lift estimates → ranked next-best variants. Dell analysts review and approve before anything ships.
What We Built: Frontend: Built the Dell Analyst Workbench on Lovable, experiment launcher, console presets aligned to treatment knobs, run history, validation/recommendation views, and a plain-language architecture walkthrough. Backend: Contributed to the FastAPI service and orchestrator with skill-based agents (validation, causal, generation, recommendation), benchmark parquet ingestion, and optional Azure OpenAI + LangSmith tracing.
Why It Matters: The design is standardized and auditable: statistics and checks stay explicit; LLMs assist with diagnostics and proposals, not opaque black-box decisions. The architecture is built to extend toward adaptive allocation (bandits/RL) without requiring it on day one.
Built as BUS AN 599 Practicum, Dell Capstone · Business Analytics & AI at the University of Washington.