Projects & Systems

ARTHASETU 2.0 is an adaptive financial inclusion platform built for BUILD BANK 2026 (Track 1: Financial Inclusion for the Underbanked). India has over 300 million credit-invisible gig workers — street vendors, domestic workers, delivery riders — who lack traditional credit histories despite having verifiable trust signals like rental payments, medical expenses, and bill payment histories. This project set out to bridge that gap by creating a system that could assess creditworthiness using alternative data, while simultaneously being accessible enough for first-time financial users who may never have interacted with a banking app before.

At the core of the platform is a statistical user-profiling engine that dynamically adapts the interface, guidance, and pacing for each user. Using K-Means clustering, the system segments users into behavioral cohorts and adjusts the onboarding flow accordingly — simpler language and larger buttons for first-time users, more advanced options for digitally literate ones. The credit scoring model combines XGBoost (achieving 99.5% AUC-ROC) with Random Forest for ensemble predictions, trained on a synthetic dataset that mimics real-world gig worker financial patterns. Feature engineering included payment regularity scores, expense volatility indices, and social trust signals derived from peer references.

Security was a first-class concern, not an afterthought. The platform implements a 10-layer post-quantum security stack: Zero-Knowledge Proofs (ZKP) allow users to prove creditworthiness without revealing raw financial data, Fully Homomorphic Encryption (FHE) enables computation on encrypted data so the server never sees plaintext, and Post-Quantum Cryptography (PQC) algorithms like CRYSTALS-Kyber and CRYSTALS-Dilithium protect against future quantum computing threats. This ensures that even if an adversary records encrypted traffic today, they cannot decrypt it once quantum computers become practical.

The project was built as a proof-of-concept for the hackathon but is designed with production in mind. The architecture separates the ML inference layer from the security layer, allowing either to be upgraded independently. Future work includes integrating with actual UPI transaction APIs, deploying the model on edge devices for offline scoring in rural areas, and adding multi-language support for Hindi, Bengali, and Tamil interfaces.

This project tackles two interconnected problems in disaster management: predicting where floods will occur and optimizing evacuation routes when they do. India is one of the most flood-prone countries in the world, with recurring monsoon disasters affecting millions across Assam, Bihar, West Bengal, and Kerala every year. The goal was to build a national-scale model that could both predict flood risk at a granular level and compute optimal evacuation strategies in real-time, covering India's complete geographic extent from 8°N to 37°N latitude and 68°E to 97°E longitude.

The prediction component uses a Random Forest classifier trained on 10,000 geospatial observations incorporating rainfall intensity, river discharge rates, elevation data from SRTM DEM, water levels, soil moisture, and population density. The dataset was validated against 5,057 historical flood occurrences sourced from the India Meteorological Department and Central Water Commission. Feature selection was done using correlation analysis and SHAP values to identify the most predictive variables. The model achieved 96.7% AUC-ROC, with the stratification division of 10.84% high-risk, 77.13% moderate-risk, and 12.03% low-risk zones — a distribution that aligns with real-world observations.

For evacuation planning, the project implements Particle Swarm Optimization (PSO) — a metaheuristic algorithm inspired by the social behavior of bird flocking. The PSO treats each evacuation route as a particle in a swarm, iteratively improving routes by balancing distance, road capacity, flood severity along the path, and vehicle availability. The fitness function incorporates real-time flood depth projections and road inundation maps. Over 200 iterations with a swarm size of 50, the algorithm converged on routes that reduced evacuation time by 33% compared to traditional shortest-path algorithms, which fail to account for dynamic flood conditions.

The entire pipeline is built in Python using GeoPandas for spatial operations, Scikit-learn for the classifier, and a custom PSO implementation optimized with NumPy vectorization. Results are visualized on interactive Folium maps that show risk zones, evacuation corridors, and shelter locations. The model is designed to be extensible — integrating live satellite imagery from Sentinel-1 SAR and real-time river gauge data from CWC could enable actual early-warning capabilities. Future work includes coupling the model with drone-based survey data for even finer resolution in urban flood-prone areas.

Pathway Track represents a paradigm in interactive machine learning pedagogy — a fully client-side, mathematically rigorous exploratory artifact designed to bridge the gap between theoretical State Space Model (SSM) dynamics and the architectural innovations introduced by Pathway's Dragon Hatchling (BDH) family. Developed for the NeurIPS 2026 Education Track, the explainer operationalizes a single falsifiable hypothesis: that fixed-size linear recurrent states, while achieving constant O(1) memory complexity per token, suffer from exponential associative memory decay governed by the spectral radius of the transition matrix $ρ(A)^L$ — a fundamental limitation that selective input-dependent gating and multi-scale Hebbian synaptic plasticity systematically resolve

Every computational substrate — from HiPPO-structured transition matrix generation and eigenvalue decomposition via 2×2 rotation block factorization to cosine-similarity fidelity probes and outer-product fast-weight updates — executes in real-time within the browser at sub-16ms latency, eliminating all precomputation, scripted animation, or server-side dependency.

The engine implements four distinct architectural paradigms for head-to-head comparative evaluation: classical linear SSM recurrence with HiPPO-inspired continuous memory initialization, selective state space modeling following the Mamba S6 discretization scheme where input-dependent gating coefficients Δ_t = softplus(w^T·x_t) modulate noise suppression, Dragon Hatchling synaptic plasticity employing the Hebbian outer-product update rule W_t = λ·W_{t-1} + η·(y_t ⊗ x_t^T) to maintain O(d²) associative capacity without sequence-length-proportional memory allocation, and BDH CQ continuous latent reasoning which unrolls recurrent micro-steps z_{τ+1} = LayerNorm(z_τ + f_θ(z_τ, c)) within the hidden manifold to achieve multi-hop deductive inference without discrete token emission