Skip to content

Research

Generative AI, machine unlearning, and responsible model editing

Selected projects, publications, and patents from my Ph.D. research and applied work — spanning foundation models, diffusion-based generation, and the tools to understand and edit what these models have learned.

Research

Research Interests

The problems I keep coming back to, across papers and production systems.

Machine Unlearning

Selectively removing the influence of specific data, concepts, or identities from trained models without costly full retraining.

Model Steering & Concept Editing

Locating and editing fine-grained concepts inside foundation models using sparse, neuron-level representations.

Mechanistic Interpretability

Understanding what is actually happening inside large models well enough to intervene on it responsibly.

Multimodal Foundation Models

Vision-language models and other multimodal systems, and how their internal representations can be understood and controlled.

Diffusion Models & Generative Video

Text-to-image and text-to-video generation — from diffusion fundamentals to structured, multi-scene video pipelines.

Large Language Models

LLM reasoning and scene-planning as a control layer for downstream generative and multimodal systems.

Selected Work

Research & Applied Projects

A mix of published research, active submissions, and the generative video system built at Valoi.

IEEE IJCB 2025

CURE: Centroid-guided Unsupervised Representation Erasure for Facial Recognition

Removes the influence of specific identities from facial recognition models through centroid-guided, unsupervised representation erasure — without full retraining.

Machine UnlearningFacial RecognitionBiometric PrivacyRepresentation Erasure

Under Review, IEEE T-BIOM

NEST: Fine-Grained Concept Unlearning via Neuron-Level Targeted Editing

Targets and edits individual neurons to perform fine-grained concept unlearning in foundation models, aiming for precise removal with minimal collateral impact on unrelated capabilities.

Neuron-Level EditingFine-Grained Concept UnlearningFoundation ModelsResponsible AI

IEEE WETICE 2026

Towards Responsible AI: Concept-Level Steering in VLMs via Sparse Mechanistic Representations

Uses sparse, mechanistic representations to identify and steer specific concepts inside vision-language models — a step toward more controllable and interpretable multimodal systems.

Sparse RepresentationsVision-Language ModelsMechanistic InterpretabilityConcept Steering

Under Review, NeurIPS 2026

Counterfactual Feature Steering for Adjacency-Preserving Identity Unlearning

Applies counterfactual feature steering with sparse autoencoders to unlearn specific identities while preserving the surrounding representation structure of a model.

Sparse AutoencodersIdentity UnlearningFeature SteeringModel Editing

Startup — Valoi

Valoi Generative Video System

An AI-powered text-to-video system that converts educational text into videos for dyslexic learners — combining LLM-driven reasoning and scene planning with diffusion-based generative video into a unified, multi-scene pipeline. Underlies two U.S. patent applications.

LLM ReasoningScene PlanningText-to-Video GenerationDiffusion ModelsMultimodal PipelinesAccessible EducationProduction AI Engineering

IEEE ICMLA 2024

Segmentation of Maya Hieroglyphs through Fine-Tuned Foundation Models

Fine-tunes foundation models for pixel-level segmentation of Maya hieroglyphs, applying modern computer vision to a cultural heritage and archaeological documentation problem.

Foundation ModelsComputer VisionCultural HeritageImage Segmentation

Research Output

Publications

Peer-reviewed and preprint work spanning machine unlearning, model steering, mechanistic interpretability, and computer vision.

  1. 01
    PublishedFirst Author

    CURE: Centroid-guided Unsupervised Representation Erasure for Facial Recognition

    IEEE IJCB 2025 · Shivam et al.

    Machine UnlearningComputer Vision
  2. 02
    PublishedFirst Author

    Segmentation of Maya Hieroglyphs through Fine-Tuned Foundation Models

    IEEE ICMLA 2024 · Shivam et al.

    Computer Vision
  3. 03
    Under ReviewFirst Author

    NEST: Fine-Grained Concept Unlearning via Neuron-Level Targeted Editing

    Under Review, IEEE Transactions on Biometrics, Behavior, and Identity Science (T-BIOM) · Shivam et al.

    Machine Unlearning
  4. 04
    PublishedFirst Author

    Towards Responsible AI: Concept-Level Steering in VLMs via Sparse Mechanistic Representations

    IEEE WETICE 2026 · Shivam et al.

    Vision-Language Models
  5. 05
    Under ReviewFirst Author

    Counterfactual Feature Steering for Adjacency-Preserving Identity Unlearning

    Under Review, NeurIPS 2026 · Shivam et al.

    Machine Unlearning
  6. 06
    Under ReviewCo-Author

    A Custom Transformer-based Model for Eye Rubbing Detection in Diagnosed Keratoconus

    Under Review, IEEE ICDH

    Healthcare AIComputer Vision
  7. 07
    Under ReviewCo-Author

    Feature Selection for Palmprint Verification Using an Improved Binary Sand Cat Swarm Optimization Algorithm

    Under Review, IEEE IJCB

    Computer Vision

Innovation

Patents

Named inventor on two U.S. patent applications involving generative video and structured multi-scene text-to-video transformation.

Text-to-Video Conversion

U.S. Patent Application Publication · Publication No. US-2025-0336199-A1

Covers the generative video system underlying Valoi's text-to-video product.

Role: Inventor

Compliance-Adaptive System and Method for Structured Multi-Scene Text-to-Generative Video Transformation

U.S. Patent Application — Under Review · Attorney Docket No. VLAI/0003USL

Covers a compliance-adaptive method for transforming text into structured, multi-scene generative video.

Role: Inventor