Projects

ComfyUI node implementing ByteDance Infinite You for identity-preserving image generation — multi-face support, gaze control, and Flux integration.

Python toolkit for building ComfyUI workflows programmatically — typed node wrappers, graph primitives, and JSON export for the ComfyUI API.

KTH DD2424 project — fine-tuned Llama-3-8B with LoRA on samhog/psychology-10k for psychological advice, compared against an LSTM baseline, and generated synthetic training data via AutoGen multi-agent simulation with GPT-3.5 Turbo.

KTH II2202 project — lung segmentation pre-processing and contrastive SSL pre-training on chest X-rays (CheXpert, VinBig) to build reusable foundation models for downstream diagnosis.

Journal of Visual Communication and Image Representation. Domain randomization for single-image dehazing with synthetic-to-real transfer.

International Journal of Imaging Systems and Technology. Self-supervised pretraining for cervical image object detection.

Top 10 in India at the Lowe's hackathon — modified A* pathfinding app to navigate customers through a store layout and optimize multi-item shopping trips.

Single-car elevator controller implemented in Verilog — floor tracking, request queueing, and door control logic.

Classic Pong implemented on an MSP430 microcontroller, rendered on an 8×8 LED matrix display.

Unity3D synthetic data generator for training vision models — keypoint detection, depth maps, object localization, and classification from procedurally spawned 3D scenes.

Extracted depth from RGB images using a GAN, trained on synthetically generated depth datasets from Unity3D.

Ball-E

An adventure game built in Unity3D, exploring physics, level design, and interactive gameplay mechanics built when I was in high school.