B.Tech CSE, Amity University Mohali

Deepinder Singh

I study how language models and AI agents behave, and build tools that make them safer and easier to understand.

  • LLM interpretability
  • AI agent security
  • Medical imaging
First-author paper · IEEE INDISCON 2026
View research Download CV GitHub

Paper accepted, OCIT 2026 · Research Intern 2026 · 200+ LeetCode problems

Currently: trust scores for multimodal activity recognition; continual learning for language models.

deepinder@ubuntu: ~ deepinder@ubuntu : ~ $ whoami deepinder deepinder@ubuntu : ~ $ cat focus.txt LLMs · AI agent security · medical imaging deepinder@ubuntu : ~ $ ls projects/ kaarigar/ fairlens/ vibeguard/ deepinder@ubuntu : ~ $ echo $STACK PyTorch · FastAPI · Next.js · MongoDB deepinder@ubuntu : ~ $ cat now.txt trust scores · continual learning deepinder@ubuntu : ~ $ █

Research

Papers and ongoing work

I research how language models and AI agents behave, and apply machine learning to medical imaging. I work carefully: patient-level data splits, data-leakage audits, and checking every reported number against the source data.

Research Intern, Amity University Mohali Jun–Jul 2026 (45 days), with research continuing since

Papers

Presented · proceedings forthcoming

Information Geometry of Prompt Quality: Representation Sensitivity as a Geometric Signal for LLM Behavior

Deepinder Singh, Abhinav Nara, Vishwanath Bijalwan*, Rajni Mohana

7th IEEE INDISCON 2026, MNIT Jaipur · Presented 12 Sept 2026

Asks whether the geometry of a language model's internal representations can signal prompt quality in a single forward pass, with no training. Tested on four open language models, eight prompt styles and 127 tasks.

Accepted

Categorical Verification Layer for Agentic AI: A Functional Calculus for Type-Safe Workflows

Gunakshh Sharma, Deepinder Singh, Vishwanath Bijalwan*

OCIT 2026 (24th OITS International Conference on Information Technology), NIT Kurukshetra · 17–19 Dec 2026

Proposes a verification layer that keeps security-critical decisions in typed code instead of trusting a language model's output, to protect AI agents from prompt-injection attacks.

* Corresponding author

Ongoing research

MRI-based glioma grading across hospitals

Classifying high- vs low-grade brain tumours from MRI across three public cohorts (BraTS2020, UCSF-PDGM, UTSW-Glioma), with patient-level splits, data-leakage audits and external validation.

TrustMI-HAR

Per-prediction trust scores for multimodal activity recognition: testing whether sensors that agree with each other can still be jointly wrong.

Selected Projects

Things I have built

Kaarigar

Aug–Sep 2026 · Smart India Hackathon 2026 · Team of six

A platform connecting skilled tradespeople (kaarigars) with customers. Separate customer and kaarigar apps are both working, with the interface in three languages. Shortlisted in the university round of Smart India Hackathon 2026.

Free hosting: the live demo may take about 30 s to wake up.

React TypeScript Vite Tailwind Express MongoDB

FairLens

Apr–May 2026 · Built for an online AI hackathon

Identifies bias in datasets and machine learning models with visual analytics and AI-driven insights.

Python FastAPI Next.js Machine learning
Read Case Study

VibeGuard AI

May 2026 · Started at a hackathon, rebuilt from scratch afterwards

AI-powered code security scanner that detects vulnerabilities in GitHub repositories and ZIP uploads with clear explanations and severity-based results.

Next.js FastAPI Python AI Security

PIXEL / SANKALP Collaboration Platform

Mar–Aug 2026 · Contribution to a team platform

A full-stack collaboration platform for managing proposals, voting, and project workflows within a developer community. I added features and fixes to the existing platform.

Next.js MongoDB Comments & replies Proposal editing Theme switcher

More on GitHub

Skills

What I use, and where

ML & research

  • Python
  • PyTorch
  • scikit-learn
  • Hugging Face Transformers

Used in: INDISCON paper, glioma research, TrustMI-HAR

Evaluation

  • Permutation tests
  • Bootstrap confidence intervals
  • Rank correlation

Used in: Paper analyses, agent-security evaluation

LLMs & agents

  • OpenAI, Anthropic and Gemini APIs
  • AgentDojo

Used in: Agent-security experiments

Medical imaging

  • nibabel (NIfTI)
  • PyRadiomics

Used in: Glioma research

Web

  • Next.js
  • React
  • TypeScript
  • FastAPI
  • Express
  • MongoDB
  • Tailwind

Used in: Kaarigar, FairLens, VibeGuard AI

Tools

  • Linux
  • CUDA GPUs
  • Git
  • LaTeX

Used in: Research work and papers

Honours & credentials

Competitions and training

  • 1st of 21 teams, Re-ignite 2.0 2K26 Ideathon

    PIXEL Club × Hrfy.ai · Feb 2026 · Idea-stage concept

    View certificate

  • Shortlisted, university round, Smart India Hackathon 2026

    Kaarigar · Aug 2026

  • AWS Academy Cloud Foundations

    Amazon Web Services · Training badge

    View badge

  • AWS Cloud Quest: Cloud Practitioner

    Amazon Web Services · Training badge

    View badge

Contact Me

Research collaborations, internships, or questions about my papers.