Esha Roy

Problem: real-time voice agents needed repeatable evaluation without rebuilding the evaluation stack for every project.

Worked on a LangGraph/Pipecat-based voice and chat agent ecosystem, and built an in-house pip-installable Python evaluation package. Teams provide a lightweight configuration plus a golden script defining expected tools, outputs and trigger stage; the library then runs standardized checks and produces terminal metrics and result files.

Implemented for LangGraph; designed so the evaluation pattern can be extended later without claiming support that does not yet exist.

Problem: identify meaningful usage patterns where labelled classes did not exist.

This was not a forecasting project. The data was time-indexed, but the modelling task was behavioural segmentation: multi-scale temporal feature engineering, dimensionality reduction with UMAP, KMeans clustering, and variance-ratio analysis to assess cluster separation.