Josh Das

Implemented an explicit Euler finite difference solver and a physics-informed neural network (PINN) for the 1D heat equation in a shared Python codebase for direct comparison. Verified 2nd order spatial convergence across 4 grid sizes, then trained the PINN to ~0.5% relative L2 error and ~0.015 maximum absolute error against the reference solver. Designed spatiotemporal error heatmaps and loss curves to diagnose sensitivity to timestep and boundary constraint weighting.

Assembled a vibration monitoring rig using a piezoelectric disc and Arduino, then built an end-to-end classification pipeline from raw sensor data. Achieved 67% test accuracy across 5 impact positions on a clamped beam (vs. 20% random baseline) using a Random Forest over a 13-feature set combining time-domain (RMS, energy) and spectral features (FFT centroid). Validated experimental data against Euler-Bernoulli beam theory, diagnosing where sensor mass loading and boundary compliance deviate from ideal analytical predictions.

Co-developing an open-source web platform that annotates and visualises predicted protein structures from DeepMind AlphaFold. Accepts UniProt IDs, protein names, FASTA sequences, and uploaded PDB/CIF files, then fetches motif and domain annotations from UniProt and InterPro, aligns annotation positions to the structure's residue numbering via SIFTS to account for sequence gaps and offsets, and renders the annotated structure in an interactive mol* 3D viewer with each region colour-coded and selectable. Currently preparing a manuscript on this work.

Architected MediVisual, a full-stack diagnostic tool for visualising breast cancer lesions on 2D patient models from ultrasonic sensor data, pitched successfully to 50+ oncologists. Built MongoDB aggregation pipelines for longitudinal patient tracking and lesion growth trend quantification. Optimised backend I/O throughput by 12% via profiling and indexing, and delivered a production-ready MVP in 4 weeks using Docker and CI/CD pipelines.