About

I'm a senior software engineer in R&D at Brac IT Services, where I work on agentic AI, large language models and retrieval-augmented generation. Before that I spent close to four years at CholPori, an online learning platform, building recommendation and context-aware content delivery systems.

In my own research I ask what a model really learns from the signal it is trained on. I implemented GRPO from scratch to see how a lenient reward grader changes what a language model does, and I built a generative recommender to test whether product images improve semantic IDs. In both, I report the results that went against my hypotheses as plainly as the ones that supported them.

I'm open to research collaborations in these areas. If you're working on something related, I'd be glad to hear about it.

Research interests

  • Reinforcement learning for language models
  • Recommender systems and generative retrieval
  • LLM agents and retrieval-augmented generation
  • Evaluation and benchmark reliability
  • Computer vision

One question runs through all of these: when a model is optimised for a proxy, what does it learn in place of the task?

News

  • May 2025 Career

    Joined Brac IT Services Ltd. as a Senior Software Engineer (R&D).

  • Sep 2021 Career

    Joined CholPori Technologies Ltd. as a Software Engineer.

  • Jul 2021 Award

    Best Employee of the Month, Infonet Associates Ltd.

  • Apr 2021 Career

    Joined Infonet Associates Ltd. as a Software Engineer.

  • 2018 Education

    Graduated from IUBAT with a B.Sc. in Computer Science and Engineering.

Selected projects

All projects →
Paralegal Chat Assitance

Odhikar

Odhikar transforms clients’ natural Bangla narratives into actionable legal case information through voice-first interaction, domain-specific LLMs, urgent-case escalation, local inference for privacy, and human review before complaint generation.

GeminiLangGraphRAGLoRAQwen2.5-0.5B
 RL fine-tuning of an LLM

GRPO from Scratch

I implemented GRPO and trained one small model on GSM8K four times, changing only the grader used as reward, then scored every model with the strict grader.

Strict reward: 9.1% to 52.3%. Lenient reward: about 0%
GRPOLoRAQwen2.5-0.5BGSM8K
 Generative recommendation

Multimodal Semantic IDs

I built a TIGER-style generative recommender and compared four ways of building item codes from text and images on Amazon Reviews 2023.

Late fusion: NDCG@10 0.0294 to 0.0332, but SASRec still wins
RQ-VAESigLIPSentence-T5SASRec
 Medical object detection

Brain Tumor Detection in MRI

I compared a one-stage, a two-stage and a transformer detector on 5,249 MRI slices under one data pipeline and one COCO evaluation.

YOLOv8m 0.781 mAP@0.5:0.95 against 0.562 and 0.508
YOLOv8Faster R-CNNDETR
 Local RAG

Smart Research Assistant

I built a fully local question-answering system over research papers, with a chunking stage that keeps reference lists out of the index.

Answers come with their source files and similarity scores
ChromaDBMiniLMLlama 3.2

Let's collaborate

I'm open to research collaborations, especially on reinforcement learning for language models, recommendation and evaluation. A short email about what you're working on is the best way to start.

Get in touch