Projects

Protein language models for antimicrobial peptide design

Antimicrobial peptides are short proteins that kill bacteria, and designing new ones by hand is slow. I fine-tuned ESM-2, a 35M-parameter protein language model, with masked language modelling so it could propose novel beta-hairpin peptides, bringing training loss from 4.93 to 1.05.

The model's suggestions then went through multi-stage biophysical filters on net charge, amphipathicity, and structural propensity, with a 91.4% validation pass rate across 180+ candidates. A good part of the work was data: cleaning contaminated structural data and generating synthetic training sequences.

ESM-2 · PyTorch · Hugging Face Transformers

IoT computer vision pipeline for automated waste segregation

A CNN-based classifier running on a physical sorting rig: a camera sees the item, the model classifies it in real time, and actuators move it to the right bin. I owned the vision model and the low-latency ingestion pipeline across sensors and microcontrollers, with fault-tolerant edge-to-cloud telemetry, working with mechanical, firmware, and software teammates.

CNNs · Vision Transformers · IoT · background removal for the Trashnet model