Every case study below answers the same four questions in the same order: the problem, the system, the decisions, the result. The third one is the only one worth your time.
Anyone can list what they shipped. What you cannot get from a repository is what was considered and rejected, so each decision is written as what I chose, what I chose against, and what that trade actually cost.
01
In production
Internal applications at Commure, running healthcare revenue-cycle operations. Described in general terms: no patient data, no identifiers, no screenshots.
On whether synthetic training data can stand in for the real thing in a text classification task, and where it stops being able to. Six authors; presented at the 26th International Conference on Computer and Information Technology.
Evaluating AI-powered Bangla SRHR chatbots using large language models
Undergraduate thesis · 2026
An evaluation methodology for Bangla LLM and RAG systems, which mostly meant confronting how little of the standard English evaluation toolkit survives the change of language.
04
Also built
Coursework and side builds. A paragraph each, because a case study on a two-week project would be padding. These have no page of their own, so open a row to see what is linked.