§ 01
I work on video understanding — mostly on what a model should hold on to from a long video, and how much of it can be worked out while the video is still playing rather than after the fact. Most of my thesis has gone into the streaming version of that, in egocentric procedural video: cooking, assembly, lab protocols, where you get one frame at a time, no future context and no second pass. That rules out most of what the field currently does well.
VidParse (ECCV 2026) does it without training anything. It segments the video using features anchored in hand–object interaction, then decodes the sequence of steps with a beam search that can only follow transitions the task graph permits. Everything runs off frozen foundation models, and on multi-step parsing it beats trained online baselines by up to 10×.
The task graph is the weak part: VidParse has to commit to one up front. What I’m working on now is using a VLM to build a hierarchical memory in a single streaming pass, structured so that different questions can be answered from it later.
Earlier work was about what learned representations already encode. Latent-INR found discriminative semantics inside implicit neural representations of video, which were supposed to be doing reconstruction. LEIA learned articulation latents that survive a change of viewpoint. CSD made style a measurable, separable property of diffusion output. TREND, from a robotics collaboration, is preference-based RL that holds up when the preferences — human or VLM — are noisy.
I spent eight years in industry before the PhD: analyst work at a few banks, then computer vision for automotive safety at Netradyne and geospatial ML at Swiggy, plus four summers at Amazon during it.
The direction I am trying to pull these threads toward is robotics and continual learning: systems that keep learning from what they watch, instead of being trained once and then frozen.
§ 02
An online, training-free framework that parses egocentric procedures by pairing manipulation-anchored features from frozen foundation models with a beam search constrained by an induced task graph — up to 10× better multi-step parsing accuracy without a single gradient update.
Preference feedback from human or VLM annotators is noisy. TREND combines few-shot expert demonstrations with a tri-teaching strategy to mitigate that noise in preference-based reinforcement learning.
Can we measure the style similarity between images? We extract style into Contrastive Style Descriptors (CSD) and use them to study style replication in image generation models.
We show semantic capabilities in Implicit Neural Representations by proposing a framework that learns discriminative semantics in videos.
Modeling unseen 3D articulation states by interpolating across a learnable, view-invariant latent embedding space.
Emergent communication via mid-level patches in a referential game played on a large-scale image dataset.
Unsupervised point-of-interest mapping — polygon boundaries recovered from GPS traces, OpenStreetMap and address information in dense urban environments.
Defining a paradigm for unbiased evaluation in open worlds.
§ 03
Detecting the driving manoeuvres that actively reduce risk, rather than only the ones that cause it, from vehicle-mounted video and telemetry analysed on-device. From my time building automotive-safety computer vision at Netradyne.
§ 04
§ 05
VidParse is accepted to ECCV 2026. Online, training-free parsing of egocentric procedures.
Named an Outstanding Reviewer at CVPR 2025.
Three papers accepted to ECCV 2024 — two of them as shared first author.
Applied Scientist Intern at Amazon Fashion, on diffusion-based controllable indoor scene creation.
Applied Scientist Intern at Amazon Fashion.
Organized the workshop Dealing with Novelty in Open Worlds (DNOW) at WACV 2023.
Applied Scientist Intern (Co-op) at Amazon, working on visual geolocalization.
Organized ObjClsDisc: In-the-Wild Object Discovery Challenge, part of the Visual Perception and Learning in an Open World workshop at CVPR 2022.
Applied Scientist Intern at Amazon — edge-based object detection models, deployed for field trials.
Organized the workshop Dealing with Novelty in Open Worlds (DNOW) at WACV 2022.
PatchGame is accepted to NeurIPS 2021.
Began my MS in Computer Science at UMD.
Joined Swiggy as a Research Intern, then Data Scientist.
Built computer vision models and other cool things at Netradyne.
§ 06