I work on NLP, LLM evaluation, and AI safety, mostly in the places mainstream benchmarks skip: low-resource languages like Bangla, figurative and legal text, and the gap between what a model understands and what it contains. I build datasets and evaluation protocols meant to falsify claims about model capability, not confirm them.
I'm also an engineer at heart: when I want to understand something, I build my own version of it, from a version control system to a code extractor in Rust.
BSc Software Engineering, Islamic University of Technology (IUT), Bangladesh, 2026, First Class with Honours · Affiliate Researcher, Systems and Software Lab.
- When Words Don't Mean What They Say: Figurative Understanding in Bengali Idioms. LREC 2026. paper · data
30 LLMs on 10,361 Bengali idioms: none exceeds 50% accuracy, against 83.4% for humans. - LLM-Ideoplasticity: Measuring Ideological Plasticity in the Political Behavior of LLMs. AACL-IJCNLP 2026. arXiv · code
- AIDG: Information Extraction and Containment Asymmetries in Multi-Turn LLM Dialogue. REALM @ EMNLP 2026. arXiv · code
- AREG: Adversarial Resource Extraction Game for Persuasion and Resistance in LLMs. MusIML @ ICML 2026. OpenReview · code
- Two Emojis of Difference: What Multilingual Affective Generation Benchmarks Actually Measure. MRL @ EMNLP 2026. arXiv
Plus 7 preprints, on literal bias in Bengali idiom understanding, temporal reasoning, cross-lingual retrieval, LLM safety in Bangla, LLM evaluation geometry, political alignment, and Bangladeshi legal text. All on Google Scholar.
- Bangla Bagdhara: 10,361 Bengali idioms under a 19-field annotation schema. The earlier public release has 5,200+ downloads.
- Bangladesh Legal Acts Dataset: 1,484 acts from 1799 to 2025, 35,633 sections, bilingual, annotated with the legal system in force at the time.
| Project | What it does | Stack |
|---|---|---|
| Reko | Extracts every function from a repo in 15 languages into a fixed IR, with byte-exact reconstruction and built-in semantic search | Rust ONNX sqlite-vec |
| ISHKAPON AI | Local-first desktop tutor: every number is computed in a sandboxed Python runtime, so the model explains results it didn't make up | Electron React Pyodide |
| Pi Daytona Chat | Coding agent where every conversation gets its own isolated sandbox, with every tool call logged | TypeScript Convex Daytona |
| AIManim | Turns a plain-language math prompt into a narrated Manim animation | Python Manim LLMs |
| Geon | Single-file DSL that renders mathematical figures to SVG, with a hand-written expression engine | JavaScript SVG |
| WhatTheHell | Turns your last failed command's error into one actionable sentence | C++17 CMake |
| IUT Fest | Registration and operations platform for a six-event university festival | Laravel MySQL Redis |