Projects

Odhikar enables legal-aid clients, including those with low literacy, to describe their situations naturally in Bangla through voice-first interaction, progressively transforming their narratives into structured, actionable case information with safety escalation for urgent cases. It combines domain-specific LLM adaptation and local inference for privacy-sensitive deployments, while ensuring human review before complaint generation.

What does a language model learn when the grader that pays its reward is more lenient than the real objective? I implemented GRPO from scratch and fine-tuned Qwen2.5-0.5B on GSM8K four times, changing only the reward grader. Trained on the strict grader, test accuracy rose from 9.1% to 52.3%. Trained on lenient graders, the model got better at the arithmetic but stopped writing the required answer tag, so its strict accuracy fell to about 0%. A small format bonus was the worst: within ten steps the model stopped reasoning and wrote only a tag with a guessed number.

A generative recommender gives every product a short code and generates the code of the next product a user will buy. I built one in the style of TIGER on Amazon Reviews 2023 and compared four ways of building the codes from product text and images. Late fusion beat text-only codes in all four settings, for example NDCG@10 from 0.0294 to 0.0332, while early fusion did not help. Two results went against me: no generative model retrieved a single cold item, and a plain SASRec fed the same embeddings beat every generative model.

I compared a one-stage detector, a two-stage detector and a transformer on 5,249 MRI slices with four tumor classes, all under one data pipeline and one COCO evaluation. YOLOv8m reached 0.781 mAP@0.5:0.95, against 0.562 for Faster R-CNN and 0.508 for DETR. Faster R-CNN is close to YOLOv8 at IoU 0.5 and far behind at IoU 0.75: it finds the tumors and draws loose boxes.

A question-answering system over research papers that runs entirely on my own machine. I built it from separate components so that I understood each stage. Academic PDFs end in long reference lists that match many queries and explain nothing, so my chunker measures citation-pattern density and keeps those chunks out of the index. Every answer lists its source files and similarity scores.

I fine-tuned a COCO-pretrained Faster R-CNN with a ResNet-50 backbone on PASCAL VOC 2012, replacing the detection head for the 20 VOC classes and adding flip, brightness and contrast augmentation.

I wrote a custom convolutional network with the Keras functional API, using batch normalisation, max-pooling and dropout, and trained it on CIFAR-10.

I trained a recurrent network to predict the next character. It then generates text in the style of its training data.

A microservices architecture I built on .NET Core for cloud-native applications.

My starting point for the mediator design pattern in ASP.NET Core, with custom middleware that handles responses in one place.

I implemented real-time push notifications in ASP.NET Core using Server-Sent Events.

I built a one-to-one chat with registration, login and message history, using Angular, ASP.NET Core, SignalR and SQL Server.