Built an AI agent workflow for healthcare referral-ranking at a 480-hacker, OpenAI-sponsored hackathon (Databricks Data + AI Summit, Apps & Agents for Good track): a Databricks App matching patients to healthcare facilities by care need and location, ranked by an embeddings-based relevance-scoring model across 10,000+ facility records. Followed Git version control, CI/CD, and LLMOps-style monitoring practices to ship a working demo end-to-end within the event window.

Built and deployed a free experiment design tool for startup PMs. Input your CVR and traffic to get sample size, runtime, CUPED variance reduction, Bayesian probability, sequential testing boundaries, live experiment analyzer, and an experiment risk score out of 100 that grades your setup across 5 dimensions before you launch.

Designed and analyzed a controlled A/B experiment identifying a statistically significant 35% engagement uplift using hypothesis testing, p-values, and confidence intervals. Defined guardrail KPIs to monitor for unintended negative effects, validated the result with a power analysis for sample sizing, and documented a clear rollout recommendation.

Cross-validated a true treatment effect estimate across three methods (propensity score matching, difference-in-differences, T-learner uplift modeling) on a non-randomized feature rollout across 100K users. Root cause analysis uncovered that the naive estimate had overstated the causal effect by 85.7%, with the uplift model built from scratch using scikit-learn. Checked that all three causal methods converged on a consistent estimate before trusting the result.