Ideas & Exploration
Exploring next-generation security standards by researching and modeling Post-Quantum Cryptography systems, specifically focusing on lattice-based mathematical frameworks to resist quantum computing attacks.
Exploring how reducing the bit-width of neural network weights reduces memory and computation cost — enabling powerful models to run on constrained hardware. The Digit Recognition project served as a testbed for binary input encoding.
Studying BinaryConnect and XNOR-Net — networks where weights and activations are constrained to {−1, +1}. XNOR and popcount operations replace multiply-accumulate, making BNNs fast on standard hardware without special accelerators.
Studying Microsoft Research's approach to extreme quantization in large language models. BitNet b1.58 constrains every weight to {−1, 0, +1} with comparable performance to full-precision models, making LLM inference viable on commodity hardware.
Studying end-to-end deployment pipelines for running quantized models locally on mobile and embedded devices — zero latency, no cloud dependency. Directly motivates the scam detector idea.
A locally-running AI system to detect scam calls in real-time on low-end Indian smartphones. Uses a quantized speech model to analyse call audio on-device — no cloud, no data sharing, instant alerts.
Extending the Digit Recognition project to a complete BNN — binary weights, activations, and inputs throughout. Goal: maximum accuracy at minimum memory footprint, deployable on ESP32.
Following the OpenCV contribution, planning to contribute to a Python ML library — NumPy, PyTorch, or Scikit-learn. Focus: numerical precision, performance optimisation, or quantization-related utilities.
If BNN-from-scratch experiments yield interesting accuracy/memory tradeoff results on constrained devices, documenting findings in a structured technical report or workshop paper.