Hi, I'm Diego Revilla.

Developed a diffusion model that estimates the complete floorplan of a house by observing a partial set of disjoint rooms, achieving 89% floorplan reconstruction accuracy when 50% of rooms are observed.

• Constructed a dual-token training pipeline for the indoor place recognition task that surpasses modern architectures such as AnyLoc & frozen DINOv3 by 12% Recall@3 on our internal dataset.
• Trained a DQN active user querying policy that exploits the model's uncertainty to choose what to ask, gaining an extra 3% accuracy over a greedy policy, while keeping prompts to less than 3.
• Built a 72 GB dataset of 10k home floorplans expressed as heterogeneous graphs, including room and pairwise connectivity semantics, classification, meshes and room-level panoramic cubemaps.

Designed and implemented a stateful reasoning architecture to iteratively solve spatially compositional VQA tasks by applying latent recursive reasoning in vision-language models without text-based CoT.

• Introduced conditioned attention shifting, which makes the model attend to different input regions at each reasoning step, reaching 80% F1 on VSR & GQA-rel (+6% over single-pass).
• Trained a specialist model that surpasses off-the-shelf foundation models InternVL2-8B & Qwen2.5-VL-7B on the same datasets with 400M parameters on relative positional and spatial tasks.
• Parallelized VLM training over 4 H100 GPUs using DDP, shrinking training time by 70%.

Conceived a progressive neural compression codec for Gaussian Splatting to enable neural rendering on bandwidth-constrained client-server infrastructures, allowing for first render times under 3 seconds.

• Reduced the total file size by modeling 3D spatial context to exploit local geometry and feature similarities to reduce uncertainty entropy, obtaining a 21% compression gain compared to HAC.
• Matched current methods in perceptual, structural and quality metrics (>27dB PSNR, >0.87 SSIM, <0.21 LPIPS) despite decreased size on the BungeeNeRF & MipNeRF 360 datasets.
• Enforced a noise-tolerant curriculum-learning strategy that increases robustness in partial scene reconstructions, recovering +2 dB PSNR compared to training without the curriculum.