Aditya Adiga

MoSSAIC explores how AI risks and failure modes change across computational substrates, architectures, and deployment contexts, arguing that safety interventions which succeed in one setting often fail when transferred to another, and that single-solution interventions break down across heterogeneous substrates. I contributed to the position paper through the development of Live Conversational Threads as a working prototype, which grounded and stress-tested its theoretical claims about substrate flexibility. With Matthew Farr, I helped scope the forward research agenda at the Groundless Residency and co-mentored the AISC 2026 MoSSAIC stream, which produced a LessWrong sequence on substrate and a literature survey. Within that stream I also co-mentored a prototype built on top of OpenEvolve, applying evolutionary search to interpretability-method discovery, which recovered weight-derived structure on toy transformers, including in a task-agnostic setting.

This project investigates interfaces for AI systems can support collective sensemaking to facilitate genuine human interaction and discernment. I designed an epistemic interface that backgrounds AI while preserving context, structure, and interpretive agency. Conversations are represented as directed acyclic graphs to maintain thematic relationships and enable non-disruptive navigation, with bookmarking and 'potential insight' marking for collaborative sensemaking. I ran multi-venue user testing across the PIBBSS Fellowship Retreat, the AI Alignment Retreat in Ooty, and the Groundless Residency, using the feedback to drive iteration, and mentored its continued development at AISC 2026 (mentee: Kseniya Parkhamchuk) toward broader deployability.

This project extends Live Conversational Threads' epistemic tooling with a pipeline that renders informal research insights through multiple formal lenses — semiformal proof, Lean 4, causal graphs, and statistical models — within a user-specified context. I mentored its development at AISC 2026, and it completed a first round of user testing with positive reception from research engineers and AI safety researchers. It's built with Next.js, TypeScript, and Tailwind CSS, backed by a Dockerised Lean 4 verifier service for machine-checkable proofs, deployed on Vercel.

This project explores MRI-based classification of Major Depressive Disorder using three-dimensional convolutional neural networks. I designed and iteratively refined model architectures and built training and inference pipelines to support systematic experimentation. I applied Grad-CAM–based visualisation techniques to analyse model activations and identify salient brain regions. The work emphasises understanding model behaviour alongside predictive performance.

Designed an OCR-based pipeline for extracting structured information from scanned invoices, using image registration, template matching, and preprocessing to support robust performance across real-world document formats.

I worked on semantic segmentation of complex tumour microenvironments in whole-slide histopathology images. Using a UNET++-based architecture, I developed models to identify tumours, stroma, and surrounding tissue under conditions of high biological variability. Collaborated closely with a pathologist to annotate underrepresented regions and iteratively refine model performance. The resulting system achieved approximately 80% intersection-over-union while remaining robust to heterogeneous tissue structure.

Building on earlier segmentation work, this project focuses on identifying tumour-infiltrating lymphocytes within inflamed stroma regions. I fine-tuned Detectron2-based instance segmentation models and integrated them with upstream tumour microenvironment predictions. Model outputs were validated with domain experts to ensure relevance for downstream immunotherapy analysis.

This project investigates regression-based biomarker scoring using multiple instance learning. I built feature extraction pipelines based on histopathology foundation models and experimented with several MIL strategies. Model outputs were evaluated not only for predictive accuracy but also for biological plausibility and clinical applicability.

I developed a parallelised pipeline for large-scale whole-slide image processing using VIPS. The system stitches approximately eight million image tiles into pyramidal TIFF images and overlays heatmaps to support interpretation of model outputs. End-to-end processing time was reduced from several hours to under ten minutes, significantly improving scalability.

This project focused on integrating AI inference pipelines with an in-house pathology image viewer. I designed workflows informed by user interviews to align model outputs with clinical interpretability requirements. API-level integration linked image storage, predictions, and visualisation layers, improving usability and enabling smoother diagnostic and research workflows.