# Rohit Kumar > Rohit Kumar builds grounded LLM systems and inference-planning tools at Armada AI. IISc AI graduate; co-first-author research at WACV 2025 and CVPR 2026 Workshop. Public portfolio summary. Use the linked pages for detailed scope and evidence. ## Main pages - [Portfolio](https://rohit.vision/) - [Résumé](https://rohit.vision/resume/) - [Résumé PDF](https://rohit.vision/assets/Rohit-Kumar-Resume.pdf) - [Engineering work](https://rohit.vision/work/) - [Sitemap](https://rohit.vision/sitemap.xml) - [Public profile JSON](https://rohit.vision/profile.json) - Contact: rohitkumar.iisc@gmail.com ## Research ### AttriStory: Fine-Grained Attribute Realization for Visual Storytelling with Diffusion Models Manogna Sreenivas, Rohit Kumar, Soma Biswas. CVPR 2026 Workshop on Generative AI for Storytelling (AISTORY). AttriStory pairs a benchmark of 200 five-scene stories across 10 artistic styles with AttriLoss, an inference-time method for binding requested attributes to the correct objects. It refines latent codes to align attribute–object attention during early diffusion steps, integrating with SDXL, ConsiStory and StoryDiffusion without retraining. With AttriLoss, StoryDiffusion VQAScore increased from 0.8363 to 0.8636 on the AttriStory benchmark (Table 1). - [Paper](https://openaccess.thecvf.com/content/CVPR2026W/AISTORY/papers/Sreenivas_AttriStory_Fine-Grained_Attribute_Realization_for_Visual_Storytelling_with_Diffusion_Models_CVPRW_2026_paper.pdf) - [Website](https://manogna-s.github.io/attristory/) ### TACLE: Task and Class-aware Exemplar-free Semi-supervised Class Incremental Learning Jayateja Kalla, Rohit Kumar, Soma Biswas. WACV 2025. TACLE learns new classes from few labels and current-task unlabeled data without retaining earlier examples. It combines task-adaptive pseudo-label selection, class-aware weighted losses and classifier alignment using labeled and confident pseudo-labeled data. Evaluated on CIFAR-10, CIFAR-100 and ImageNet-Subset100, including one-shot and imbalanced settings. On CIFAR-100 with 0.8% labels over 10 tasks, TACLE reached 92.35% average incremental accuracy versus 88.23% for SLCA+FT (+4.12 percentage points; same ImageNet-pretrained ViT, Table 1). - [Paper](https://openaccess.thecvf.com/content/WACV2025/papers/Kalla_TACLE_Task_and_Class-Aware_Exemplar-Free_Semi-Supervised_Class_Incremental_Learning_WACV_2025_paper.pdf) - [Code](https://github.com/rokmr/TACLE) - [Website](https://rohit.vision/tacle/) ## Engineering work ### [Support RAG Agent](https://rohit.vision/work/agentic-rag/) Armada AI A grounded question-answering system for Armada’s internal Customer Support team, searching runbooks, help-center content and vendor documentation. Sole owner, from discussions with the support team and an empty repository through production integration. Reduced cost per query/run to approximately one-third of the initial graph, comparing similar workloads. The shared graph node removes repeated reformulation and embedding work. Incremental sparse indexing and the common ingestion representation make knowledge-base changes easier to maintain. ### [Almanac: air-gapped document Q&A](https://rohit.vision/work/almanac/) Armada AI An on-device RAG application designed for document question answering on Armada edge hardware without internet access. Sole engineering owner across document ingestion, local model serving, retrieval and cited question answering. Document ingestion, question answering, evaluation tooling and inference planning work in an isolated VM. A fresh installation with the internet disconnected has not yet been tested. Retrieval and generated-answer quality evaluation is underway; completed answer-quality results are not yet reported. ### [LLM Inference Planning & Optimization](https://rohit.vision/work/inference-planner/) Armada AI Engine-independent predictions of KV-cache capacity, latency, throughput and serving cost from spec sheets, with optional GPU-profile calibration. Built the planner and used its predictions to choose serving configurations, then deployed them. Spec-sheet KV-cache capacity predictions were exact on 25/25 dense and MoE configurations. KV-cache capacity exact on 25/25 dense/MoE configurations without profile data. Dense TTFT MAPE 15.9% to 3.6% (n=174), batched TPOT 13.8% to 5.2% (n=242) after GPU profiling; profiled errors partly in-sample, not held-out. Throughput prediction is the weakest measured area; the 256-user evaluation has only 3 to 6 samples per architecture family. Serving-cost estimates inherit throughput uncertainty. KV-cache capacity predictions from spec sheets were exact on 25/25 dense and MoE configurations (15 dense, 10 MoE). For hybrid architectures, maximum concurrency (N_max) was within 2% on 22 of 24 configurations. Neither result uses GPU-profile calibration. For dense models, spec-sheet-only versus GPU-profiled mean absolute percentage error (MAPE) was 15.9% versus 3.6% for time to first token (TTFT; n=174), 13.8% versus 5.2% for batched time per output token (TPOT; n=242), and 7.2% versus 4.4% for single-user TPOT (n=134). Spec-sheet-only predictions require no GPU access. Profiling takes roughly a few minutes on the target GPU and measures the gap between spec-sheet and actual hardware behavior. Some evaluation runs were also used to calibrate the GPU profile, so profiled errors are partly in-sample. They do not establish held-out generalization. MAPE is an average, not a worst-case bound; individual errors can exceed it. Profiling does not improve every metric or architecture. Throughput prediction is the weakest measured area. The 256-user evaluation has only 3 to 6 samples per architecture family, so throughput accuracy is not yet well established. Serving-cost estimates also inherit uncertainty from predicted throughput. ### [AI Assistant retrieval improvements](https://rohit.vision/work/ai-assistant/) Armada AI Contributions to an inherited RAG application, distinct from the Support Agent I later owned. Contributor. Redesigned the LangGraph workflow and improved context preparation, retrieval and metadata handling. The workflow changes were adopted and deployed. Embedding-service work was completed but not deployed. ### [Diffusion-based product imagery](https://rohit.vision/work/lifestyle-generation/) Avataar AI A lifestyle-image generation pipeline that preserves the product while adapting its appearance to a new scene. Built the Flux and ControlNet pipeline; modified diffusion sampling and product-relighting work were both adopted into Avataar’s production pipeline. Modified diffusion-sampling work and product-relighting work both adopted into Avataar’s production pipeline. Qualitatively improved product shape/consistency and lighting consistency; no standardized metric supplied. ### [3D reconstruction & novel-view synthesis](https://rohit.vision/work/novel-view-synthesis/) Avataar AI Sparse-image-conditioned novel-view synthesis for 3D reconstruction at Avataar AI. Fine-tuned sparse-image-conditioned diffusion models and worked on conditioning, distributed training and SLURM, building on a senior engineer’s initial setup. ### [Cricket-Shot Predictor](https://rohit.vision/work/cricket-shot/) Independent project Video classification for cricket shots, from frame extraction to a runnable prediction demo. Built embedding-based video classifiers and fine-tuned VideoMAE. ### [Document-model evaluation pipeline](https://rohit.vision/work/model-evaluation/) Evaluation tooling A pipeline for comparing and substituting document-processing and retrieval models to select suitable components. Evaluated table-structure models, OCR models, layout detectors and embedding models, and built a pipeline that supports switching between candidates.