INDEPENDENT RESEARCH

STATISTICS / AGENTIC AI / MATERIALS

GITHUB.COM/SHARATHK4

BML MUNJAL UNIVERSITY

2026

The State of the Art Papers.

RESEARCH ANNOUNCEMENT

A new study and model in materials science is underway.

It grows out of the TiNi force-statistics work — from molecular-dynamics force traces to the energy landscape of a martensitic transition, with inference and a forward model built to be checked. Early results and code will appear here first.

Follow the work on GitHub

MUST READ ✦ ECONOMETRIC THEORY

Local-Time Limits and Calibration Limits of Spatial-Sign Delay Statistics

Overlapping time-series windows, each divided by its length, become directions, summarised by a spatial-sign matrix and the entropy gap between two such matrices. For integrated Gaussian random walks with a fixed window length, the fluctuations are governed by Brownian local time — proved while accounting for the dependence created by overlapping windows and the cancellation of signed tails. When the initial level dominates, the statistic shrinks at a different rate and its entropy gap collapses. The consequence is a calibration limit: a lower-tail rule with a fixed positive cutoff at the square-root sample-size scale can reject with probability approaching one for some non-cointegrated Gaussian pairs.

SUBMITTED TO ECONOMETRIC THEORY · 2026
Simulation figure: slow convergence of the entropy gap, and the defect under large starting levels
Independent Gaussian walks, 1,000 replications per cell.

Why this paper matters.

Spatial-sign summaries look like an easy screen for cointegration. The limits proved here explain why a fixed lower-tail rule cannot be trusted: for some non-cointegrated Gaussian pairs it rejects with probability approaching one.

Sai Sharath Chandra Koppu · Luckshay Batra — Submitted to Econometric Theory · 2026

Selected papers.

02AGENTIC AI

NetSurf: Intent-in-the-Loop Agentic AI for Web Browsing

NetSurf's new-tab page: one omnibox for a URL, a search, or a task

A locally hosted agentic desktop browser that closes the gap between high-level intent and low-level browser mechanics without giving up privacy or interactivity. Built on Electron with the Eko agent library, it runs a goal-oriented LLM reasoning loop inside the browser process itself — no cloud execution cluster, no background daemon, no extension sandbox. Physical event synthesis turns plans into coordinate-level input through the Chrome DevTools Protocol, and the reasoning stays visible to the user, with mid-task correction. On a 50-task WebArena subset it completes 76% of tasks, 24 points above a deterministic Playwright baseline, in roughly 350 MB of memory.

Kritavya Patel · Yugen Jarwal · Sai Sharath Chandra Koppu · Kiran Sharma — Research prototype

03MATERIALS SCIENCE

Martensitic-Transition Force Statistics: Inference & PDE Extension

A reconstruction and extension of the force-statistics workflow of Maitra & Gururajan (Scientific Reports, 2025). One hundred molecular-dynamics force traces of a TiNi single crystal under tension are loaded, validated and averaged; U‡, ξ‡ and Γ₀ are inferred by maximum likelihood with profile likelihoods, bootstrap uncertainty and synthetic recovery. A forward Smoluchowski–Fokker–Planck solver — finite volume, Scharfetter–Gummel fluxes, backward Euler — is checked for positivity, conservation, first-passage behaviour and grid and time convergence, then extended with curve-transition analysis and 16-setting stability sweeps across 296 runs. 201 automated tests.

Sai Sharath Chandra Koppu — Reconstruction & extension

Research — Sharath Koppu