Rohit Kumar

Trivandrum, India

Manager: Naveen Nair

• Sole owner of the Support RAG Agent, built from an empty repository and deployed for Armada’s internal Customer Support team
• Reduced Support Agent cost per query/run to approximately one-third of the initial graph on similar workloads by sharing reformulation and embedding across retrieval branches
• Built an LLM inference planner with spec-sheet predictions and optional GPU-profile calibration; used its predictions to choose serving configurations, then deployed them
• Sole engineering owner of Almanac’s local document-Q&A application, including model serving, ingestion, retrieval and SFT/LoRA/QLoRA fine-tuning
• Enabled incremental sparse-index updates and removal with online-learning BM25, addressing organizational vocabulary without full re-indexing
• Established retrieval-stage CRAG gating and end-to-end RAGAS evaluation against curated ground truth

[Support Agent] [Almanac] [Inference planning]

Bangalore, India

• Teaching assistant for Digital Image Processing and Signal System in Practice
• Worked on continual learning (L2P, DualPrompt), self-supervised learning (MoCo, SimCLR), frequency-domain filtering and image features

LLM Inference Planning & OptimizationCase study

Engine-independent prediction of KV-cache capacity, latency, throughput and cost for dense, MoE and hybrid architectures. KV-cache capacity exact on 25/25 dense and MoE configurations. Spec-sheet-only predictions need no GPU access. With brief GPU profiling, dense mean error falls from 15.9% to 3.6% for TTFT (n=174) and 13.8% to 5.2% for batched TPOT (n=242). Errors are mean absolute percentages; profiled results are partly in-sample.

Document-model evaluationCase study

Compared table-structure, OCR, layout and embedding models; built interchangeable pipeline components for model selection.