My Work

Context-preserving inpainting pipeline for large-scale privacy-preserving facial dataset augmentation

• First co-author; built a context-preserving inpainting pipeline enabling large-scale privacy-preserving facial dataset augmentation
• Presented at ICML 2025 workshops: Data in Generative Models (DIG-BUG) & WiML

Studying failure modes in MoE routers as toy models of latent misalignment

• Studying failure modes in MoE routers as toy models of latent misalignment, focusing on hidden misaligned experts and deceptive routing modes

Auditable pre-label certify-or-abstain evals under covariate shift using PSIS-κ/ESS diagnostics

• Sole author; protocol for auditable, pre-label certify-or-abstain evals under covariate shift using PSIS-κ/ESS diagnostics and self-normalised importance sampling
• Evaluated on 19 text, graph, and molecular benchmarks: ~3–13× reduction in test universes at fixed PPV floors

Production eval pipeline for RAG workflows , Affinda Group

• Delivered Azure OpenAI + RAG workflow automation integrated with SharePoint knowledge stores and CRM for a member-based enterprise serving ~1.7M customers/year
• Architected an AI-assisted intake and fraud-screening workflow for an ops platform handling ~1K inbound cases/month

Second post to verify listing order + related posts.

A tiny post to verify `/writing` and `/writing/[slug]` rendering.