Scalaris: A Scalable Enterprise Memory Architecture for Conversational AI Assistants
Develops a unified long- and short-term memory architecture for accurate, token-efficient retrieval in enterprise conversational systems.
Questions I keep coming back to
How can learning systems recognize what they know, discover what they don’t, and remain useful when the world changes?
Recognizing known concepts while discovering new categories in unfamiliar domains. I work on representations and adaptation strategies that connect domain generalization with generalized category discovery.
DG-GCD ↗ HiDISC ↗Hierarchies are everywhere—from geographic scenes to visual categories. I explore hyperbolic representations that make room for these structures, including open-world remote sensing.
HElIOS ↗How we represent language affects what models retain and how they behave. My work spans morphology-aware Indic tokenization, multilingual interventions, and medical question answering across text and images.
SuTRA ↗ Beyond Accuracy ↗At Motilal Oswal, I study structured financial generation, time-series representations, and the implications of agents sharing foundation models. Only public-safe research details are shared here.
CandleDiff ↗ Many Agents, Few Minds ↗In a Euclidean plane, circles grow linearly with radius. In a hyperbolic plane, their circumference grows exponentially at large distances—one reason to explore this geometry for branching hierarchies.
Current questions and collaborations. Submission venues, review details, and private manuscript links are intentionally omitted.
Develops a unified long- and short-term memory architecture for accurate, token-efficient retrieval in enterprise conversational systems.
Benchmarks zero-shot, last-block, and full fine-tuning strategies for time-series foundation models across Indian and US equity data.
Studies known- and novel-category discovery in unseen domains when source data remains decentralized and non-IID.
Connects fine-grained recognition with open-world discovery under cross-domain distribution shifts.
Develops a unified perspective on hyperbolic learning for hierarchical and open-world Earth observation.
A retrieval-augmented support system built around product manuals and a QLoRA-tuned language model.
Text-conditioned 2D gesture generation using MediaPipe keypoints and a GRU pose model.
A scale-recurrent network with spatial-asymmetric attention for multimodal medical images.
Vision Transformer experiments on CIFAR-100 with substantial label noise.