Publications & Preprints

Research archive

We investigate inverse rendering, environmental illumination synthesis, and market mechanism design. Our foundational discoveries power our flagship platform, Toppins.

Notice: Scientific preprints reflect theoretical inquiry. Commercial production models, neural synthesis weights, and pipelines in Toppins are strictly confidential and protected intellectual property.

6 papers

OCT 2026

Photorealistic Virtual Product Placement via Inverse Rendering and Shadow Synthesis

K. S. Narayanan, Dr. Elena Rostova • Vision Pod

Estimating ambient illumination, shadow penumbras, and geometric contact planes to integrate 3D brand assets into uncalibrated creator photographs.

Vision →

We present an inverse rendering framework that estimates depth maps, surface normals, and spherical harmonic lighting from single-view photographs. By computing exact contact occlusions and subsurface scattering, our model achieves photorealistic insertion of branded physical objects without altering the creator's original composition or aesthetic.

SEP 2026

Democratizing Brand Sponsorships: Mechanism Design for Micro-Influencer Marketplaces

Marcus Vance, Priya Raman • Algorithmic Economics Pod

Empirical analysis of an 80/20 revenue allocation model across creator tiers (4k–9k followers), proving increased campaign ROI and sustainable creator monetisation.

Economics →

Traditional influencer sponsorship structures incur prohibitive transaction costs for creators with under 10,000 followers. We formulate an automated pooling mechanism with 80% direct revenue distribution, demonstrating that micro-influencer content delivers up to 3.8x higher contextual engagement per rupee spent.

AUG 2026

Monosemantic Feature Decomposition in Sparse Multi-Modal Latent Vectors

Dr. Aris Thorne, Julian Moreau • Mechanistic Lab

Using dictionary learning with L1 sparsity penalties to isolate individual polysemantic concepts inside multi-modal models for contextual brand safety.

Interpretability →

Superposition allows deep neural networks to represent vastly more features than dimensions, making standard weights uninterpretable. We train massive sparse autoencoders across intermediate residual activations, extracting over 16 million monosemantic feature vectors for precise brand-to-image semantic matching.

JUL 2026

Single-View Depth and Environmental Illumination Estimation for Edge Devices

Dr. Wei Zhao, Mateo Fernandez • High-Performance Vision Pod

Fast monocular metric depth recovery and spherical harmonic lighting decomposition, running at 45 fps on client devices.

Vision →

Extracting accurate metric depth from uncalibrated consumer photographs allows virtual products to obey real-world physical boundaries, occlusions, and table planes. We demonstrate a lightweight vision transformer that computes surface normals and environmental maps in sub-50ms.

MAY 2026

Contextual Inpainting and Non-Disruptive Advertising: An Empirical Perception Study

K. S. Narayanan, Dr. Lucas Vogt • Cognitive Perception Pod

Tracking viewer sentiment and eye fixation: virtual product placements retain 94% organic engagement compared to 41% for traditional promotional posts.

Perception →

Eye-tracking and sentiment analysis across 5,000 digital consumers show that dedicated commercial posts trigger cognitive banner blindness. In contrast, naturally integrated virtual products maintain audience goodwill, high brand recall, and authentic aesthetic appreciation.

NOV 2025

Automated Advertising Disclosure Verification in Creator Content

Marcus Vance, Dr. Elena Rostova • Governance Pod

Developing programmatic safeguards ensuring all virtual placements comply with ASCI, FTC, and global commercial transparency guidelines.

Ethics →

Commercial integrity requires transparency. We design an automated compliance layer that embeds verifiable sponsorship metadata and ensures clear hashtag disclosures (#ad, #sponsored) are maintained across all participating creator posts without degrading the aesthetic experience.