Vikram Kaviya

I architected an end-to-end Elasticsearch indexing and backfill pipeline designed to handle over 30 million product images for a major e-commerce platform. The primary engineering challenge involved severe file I/O bottlenecks and system-breaking hierarchies during large-scale embedding generation. To resolve this, I implemented WebDataset, ensuring massive-scale system stability and optimized image reading. Because a live-update streaming system was unavailable during the initial phases, I engineered a highly optimized backfill system utilizing diff-based snapshot comparisons.

This pipeline applied targeted conditional updates using Painless scripts, serving as the primary index synchronization mechanism to maintain data freshness. On the inference side, I optimized Vision-Language Model (VLM) deployments using AWQ quantization to automate high-throughput image description generation. Finally, I built a robust "VLM-as-a-judge" evaluation framework to continuously monitor and score the retrieval accuracy of the visual search system.

I architected a modular agentic product onboarding engine and a custom, provider-agnostic orchestration SDK from first principles to standardize the LLM lifecycle. The framework was designed across three distinct layers. First, I built the foundational SDK using strict Python class schemas for state management, structured output tool registries, and multi-part messages. This layer executes stateful tool loops for multi-turn conversations until exhaustion and leverages parameter-aware deep caching via prompt and bytecode fingerprinting. Second, I engineered a step and pipeline framework that chains multiple LLM calls. Each node is a standalone Python class that seamlessly integrates probabilistic LLM reasoning with deterministic Python logic, enabling independent development, isolated testing, and dynamic task reconfiguration. Third, I implemented trace-level explainability through granular logging to track state mutations step by step. I utilized this architecture to solve complex eCommerce category standardization, developing a normalization pipeline that refined statistical clusters into canonical attributes, achieving over 85% precision in attribute mining via semantic validation.

I developed a highly scalable, multi-class classification system designed to categorize eCommerce products across an extreme label space of over 100,000 distinct categories. To manage this massive active-label space and mitigate catastrophic forgetting during model updates, I engineered a compute-efficient incremental learning pipeline that leverages knowledge distillation and a custom dual-loss mechanism combining KL-Divergence and Cross-Entropy. This architecture successfully maintained over 90% accuracy during rigorous client audits.

I subsequently extended the framework to process structured metadata (titles and specifications) by building a two-stage retrieval and LLM reranking architecture. This system ensembles three specialized retrievers to heavily prune the search space before passing candidates to a Qwen-V3 model for final refinement. By integrating hierarchy-aware logic into the pipeline, the system achieved a 50% reduction in gross search irrelevance.

I engineered an advanced auto-regressive search query corrector for eCommerce search, implementing a highly specialized two-stage training strategy. To capture complex spelling and phonetic variations beyond simple character swaps, I engineered a custom dynamic error-simulation noise engine for synthetic pre-training. This engine modeled complex, on-the-fly phonetic and segmentation errors, boosting spelling correction accuracy by 6% to 7% when fine-tuned on historical queries for domain adaptation. Furthermore, I architected a state-of-the-art Generative Reward Policy Optimization (GRPO) latent reasoning pipeline utilizing Phi-4 Supervised Fine-Tuning (SFT) distillation. To optimize the reasoning process, I designed custom dense partial rewards, specifically focusing on trace quality and diff-zone metrics, which provided continuous, directional learning signals to the model, overcoming the limitations of traditional sparse correctness metrics.