Earlier work
Course projects, competitions and industry work from before and early in my PhD.
- 2024
LLM as a discriminator for machine-generated news headlines
WPI course project
Fine-tuned a GPT model to tell human-written from generated headlines, as an evaluation signal that does not depend on BLEU or ROUGE.
- 2024
Explaining real and fake images with a vision-language model
WPI course project
Built a dataset of GPT-written explanations for GAN and diffusion fakes and fine-tuned LLaVA on it.
- 2023
Single-pass machine unlearning for text classification
WPI course project
Used error-maximizing noise to remove whole classes from DBPedia classifiers in one epoch while keeping accuracy on the rest.
- 2023
Checking deep unlearning with saliency maps
WPI course project
Grad-CAM and HiResCAM maps showed that unlearned CIFAR-10 classes lose their saliency, supporting the Fast Unlearning paper's claims, while EigenCAM's principal components stayed non-zero.
- 2021
Telephony conversational AI for financial services
Hishab
Voice banking for Bank Asia, due reporting for Hitachi India, and product ordering for Indian small businesses, built on Kaldi and NeMo ASR, TTS and Rasa.
- 2021
Car interior image enhancement
Chowa Giken, Japan
Removed glare, haze and reflections from car interior photos, using ERRNet models trained on real and synthetic data within an inference-time limit.
- 2020
Real-time vehicle tracking with fisheye cameras
IEEE VIP Cup 2020, 5th place
A modified YOLOv4 detector with an EfficientNet-B2 backbone, robust to day and night lighting and fisheye distortion.
- 2020
Unsupervised anomaly detection for drones
IEEE SP Cup 2020, 4th place
Explained abnormal accelerations and rotations in flight logs with an LSTM and attention model trained only on normal flights.
- 2020
Audiovisual emotion recognition
B.Sc. thesis, BUET
A transformer over mel spectrogram and OpenFace facial features.
- 2019
Activity recognition and privacy protection from body cameras
IEEE VIP Cup 2019, 2nd place
Classified office activities with a two-stage classifier and blurred faces, screens and washroom scenes in real time.
- 2019
Drone-embedded sound source localization
IEEE SP Cup 2019
Estimated azimuth and elevation of sound sources from a drone microphone array; became the DOANet journal paper.