Abhinaash Tiwari
Computational Mathematics · Optimization · Applied Analysis
- Kathmandu, Nepal
- abhinaash.tiwari@gmail.com
- github.com/abhinaash-broski
Education
University of Potsdam, Germany
From Winter 2026Kathmandu University, Dhulikhel, Nepal
Dec 2021 – May 2026- Thesis
- LSTM-Driven Portfolio Optimization on the Nepal Stock Exchange (NEPSE) using Differential Evolution
- Supervisor
- Dr. Khim Bahadur Khattri
Research and Conferences
Portfolio Optimization in an Emerging Market: A Differential Evolution Approach with LSTM-Based Return Forecasting
May 2026Selected Projects
LSTM-Driven Portfolio Optimization on NEPSE using Differential Evolution
Nov 2025 – May 2026- Formulated a mixed-integer nonlinear portfolio optimization problem under real-world market constraints and solved it with Differential Evolution, using a custom decode-and-repair operator to keep integer-lot solutions feasible.
- Combined sector-level LSTM networks with attention for return signals, Ledoit–Wolf shrinkage for covariance estimation, and a penalty-augmented Sharpe ratio objective; ran a systematic sensitivity analysis over the key hyperparameters.
Image Compression Using Singular Value Decomposition
May 2025 – Jul 2025- Applied truncated SVD to lossy greyscale image compression, grounded in the Eckart–Young–Mirsky theorem, and verified its error formula numerically.
- Built a Python (NumPy) compression module and command-line tool with unit tests, choosing the rank by energy retained and PSNR. A 360×360 image is stored with 9× fewer values at 33.5 dB.
- Analysed when the method fails (random noise vs. structured patterns) and compared it with JPEG.
Physics-Informed Neural Networks vs. Finite Differences for the Heat Equation
Dec 2024 – Apr 2025- Benchmarked PINNs against explicit, implicit and Crank–Nicolson finite-difference schemes for the 1D, 2D and 3D heat equations, comparing accuracy, runtime and memory under coarse discretization.
- Trained PINNs with adaptive GradNorm loss weighting, Latin Hypercube Sampling and Adam/L-BFGS. Finite differences were more accurate; PINNs were faster in 3D.
Workshops and Academic Schools
ANAIS — Annual Nepal AI School (NAAMII)
Dec 2025 – Jan 2026AESIM 2025 — Cryptography for the 21st Century
Aug 2025Leadership and Service
Vice President, Kathmandu University Computational Mathematics Club (KUCMC)
Nov 2023 – Jan 2025- Organized a national mathematics Olympiad whose problems feed into Nepal's IMO qualification pathway.
- Coordinated a mathematics hackathon, academic workshops and a national academic conference.
Skills
- Programming
- Python, C, C++, R, MATLAB
- Tools
- Gurobi, LaTeX, Git
- Languages
- Nepali (native), English (C1)
- Interests
- Optimal transport, geometric deep learning, optimization under uncertainty, stochastic modeling
Reference
Dr. Khim Bahadur Khattri
Associate Professor, Kathmandu University
khimkhattri@ku.edu.np