// 03: projects

Projects

Everything here is a public repository. Each one opens on GitHub, research pipelines, applied AI work, and the systems coursework underneath it.

github.com/kinjalumass
// publication
arXiv preprint · arXiv:2601.05076

Chain-of-Sanitized-Thoughts

Plugging PII leakage in the chain-of-thought of large reasoning models

Arghyadeep Das, Sai Sreenivas Chintha, Rishiraj Girmal, Kinjal Pandey, Sharvi Endait

A reasoning model can answer carefully and still give the private data away in its working. The chain-of-thought is treated as scratch space, so it leaks the personal information the final answer was written to protect. This work sanitizes the reasoning trace itself, without breaking the reasoning that depends on it.

  • Contributed to PII-CoT-Bench, the benchmark used to measure how much personal information a reasoning trace exposes.
  • The same question approached from the defense side that MediaTagger approaches from the attack side: what a model reveals about data it was never meant to repeat.
  • Privacy
  • LLM reasoning
  • PII
  • Benchmarking
Read on arXiv
// flagship
Flagship · 5 case studies

Boston Cyclists Union: graph analysis

kinjalumass/BCU-Graph-Analysis

Data science for safer and more equitable biking infrastructure in Boston, built with the Boston Cyclists Union through the 2026 Data Science for the Common Good program. Graph analysis over the city road and cycling network, turning urban mobility and road-safety data into evidence that holds up in a policy conversation. Engineered as a reproducible project rather than a notebook dump: a DVC pipeline, a pytest suite, GitHub Actions running tests on every pull request, and Sphinx documentation.

  • Python
  • Graph analysis
  • DVC
  • pytest
  • CI
  • Civic data
Read the case studies
// selected
// more repositories