Research Portfolio · AI Safety & Systems

Vishwa-
jeet
Adkine

Builder, researcher — building AI systems that serve underserved communities across that are envisioned to serve across the world.

Scroll
6
Research Papers
4
Deployed Systems
2
Interactive Papers
17
Age at Build Time

About

Building AI that
matters

I'm a 17-year-old builder and researcher focused on applied AI systems that address real-world gaps — from crop disease detection for Indian farmers to AI safety mechanisms that defend against prompt injection attacks. My work spans the full stack: mobile apps, browser automation, multi-agent platforms, and theoretical ML research.

Current focus: BSc Artificial Intelligence applicant at for 2027 MBZUAI, building the Stremini AI ecosystem — a cross-platform multi-agent workspace, educational platform, and AI-powered keyboard — while publishing research in spectral properties of attention matrices and gradient descent dynamics.

AI Safety Multi-Agent Systems Transformer Theory AgriTech Cloudflare Workers Kotlin / Flutter South Asia Impact

Case Studies & Investigations

All papers →
01 — Security
AI Safety

Protecting AI Chatbots from Prompt Injection Attacks

Pattern-based detection architecture with regex matching, input sanitization, and layered system-prompt hardening. Built and deployed in production for Stremini AI.

02 — Security
AI Security

AI-Driven Scam Detection via Text & Link Patterns

Multi-signal detection combining NLP text analysis, WHOIS domain age verification, and Google Safe Browsing API. Deployed as a live website trust-score endpoint.

03 — Systems
RAG / Grounding

Self-Verification via Real-Time Web Search Grounding

How I eliminated AI hallucinations on current events by injecting verified Serper search results into the system prompt before inference. ~95% accuracy on time-sensitive queries.

04 — Systems
Temporal Awareness

Solving the Knowledge Cutoff Problem in Stremini AI

Keyword detection, category-aware search query construction, current date context injection, and trusted source prioritisation — bridging the gap between training cutoff and real-world deployment.

05 — Impact
AgriTech

Problems Faced by Indian Farmers & Digital Solutions

Research into market exploitation, climate uncertainty, and resource inaccessibility facing small farmers across India. Proposes an integrated digital framework — the foundation for KhetiAI.

06 — Ongoing
AI Robustness

Malicious URL Detection with ML Classification

Ongoing research into feature engineering for URL-based phishing detection — combining lexical features, WHOIS signals, and ensemble classifiers for high-precision real-time filtering.

Theoretical ML — Made Visual

01
Optimization Theory Interactive · 4 Live Figures

Gradient Descent as a Dynamical System: Stability, Phase Transitions & Convergence

Treats each gradient update step as a discrete-time map Φ_η : ℝⁿ → ℝⁿ and derives exact stability conditions via spectral radius analysis. Characterises three regimes — geometric convergence, oscillatory divergence, explosion — with interactive bifurcation diagrams and theory-vs-empirical verification.

Open Paper →
02
Transformer Theory Interactive · 5 Live Figures

Spectral Properties of Attention Matrices in Transformer Models

Investigates eigenvalue concentration, power-law singular value decay (σᵢ ≈ C/i^α, α ∈ [0.5, 1.0]), norm scaling, and the sparsity–spectrum duality in transformer attention. Derives expressivity bounds and practical sparsity guidelines from first principles.

Open Paper →
03
AI Safety Research Flagship Research Project

Hybrid Interpretable Scam Detection for Real-World AI Safety

Full flagship research manuscript covering a unified ensemble of supervised learning, LLM semantic analysis, and rule-based heuristics with explainability and deployment realism.

Open Paper →
MBZUAI
BSc Artificial Intelligence Applicant · 2025
Stanford University
Machine Learning Certification
DeepLearning.AI
Deep Learning Specialization
Stremini AI
Founder & Lead Engineer · Cross-platform AI Ecosystem

Deployed Systems

Full portfolio →
AgriTech

KhetiAI / KhetiPlus

AI-powered crop disease detection and market transparency platform for Indian farmers. Multilingual voice interface, equipment rental marketplace, direct-to-buyer connections.

Crop Disease CV Hindi-first UI
SaaS · Restaurant Management

OrbitDine

Restaurant management SaaS with Supabase backend, RLS policy consolidation, and multi-tenant auth. React + TypeScript + Vite stack.

React Supabase Vite

Get in Touch

Let's build
something together

Open to research collaborations, AI safety discussions, internship opportunities, and conversations about building technology for underserved communities.

Context & Background

Current Status

BSc AI applicant at MBZUAI. Building Stremini AI ecosystem full-time while publishing research papers.

Research Interests

AI safety engineering, spectral properties of transformers, multi-agent architectures, agricultural AI, and prompt injection defense mechanisms.

Collaboration Areas

AI for social impact in South Asia, educational AI platforms, AI safety tooling, applied ML research with deployable outcomes.

Technical Stack

Cloudflare Workers · React · Kotlin · Flutter · Electron · Playwright · Supabase · MBZUAI K2-Think-v2 · Groq Llama 4