Meet Vaddoriya

− Part of Splunk ITSI (IT Service intelligence platform). Working on integrating Language models in the service workflow for providing summaries on group of alerts / incidents, and also analyzing the root cause of the problems / alerts occured in IT systems. − Part on Pwny-GPT project, which is a chat-bot assitant for folks using splunk products, This project leverages RAG, where in I worked on the Retrival and Re-Ranking side of the project for finding the best source of documents to be given to the Language model. − Contributed to the development of the Cisco Time Series Model (arXiv:2511.19841), an open-weight 500M parameter univariate zero-shot forecasting foundation model. The model introduces a novel multiresolution architecture extending TimesFM to accept dual-resolution inputs (coarse 1-hour + fine 1-minute), enabling capture of both long-term seasonal patterns and short-term dynamics. Trained on 300B+ data points (50%+ from Splunk Observability Cloud), achieved SOTA performance on observability benchmarks—reducing MAE from 0.627 (TimesFM 2.5) to 0.479 on 1-minute resolution forecasting—while maintaining competitive GIFT-Eval benchmark scores. − Designed and built Splunk’s AI Orchestrator Service, a LangGraph-based agentic backend for orchestrating LLM interactions with enterprise tools via MCP. The architecture features dynamic tool schema discovery, structured JSON execution planning with dependency resolution and confidence traces, and pluggable model connectors (OpenAI/Azure, Ray Serve). Integrated Langfuse for production-grade observability of orchestration runs. Deployed across Kubernetes clusters, enabling unified AI agent capabilities for Splunk’s AI-powered products while reducing feature integration time.

− Trained multi class sentence classifier for segregating sentences into different buckets inorder to input the NER the most appropriate sentences, achieved an ROC-AUC score of 89% on average for every class. − Worked in designing and developing the Recommender systems for recommending the best experts who would most likely be able to solve the business problem, have explored and researched on Neural Collaborative Filtering (NeuMF) , Factorisation Machines, Sequential Recommender Systems etc, for designing these systems, and achieved the highest hit-ratio of 0.7 on average for all industries like Sales, Finance etc, for the BERT4Rec model. − Worked in developing and designing Learning to Rank (LTR) models inorder to re-rank the experts which are best suited for the given business problems. − Worked on developing custom information retrival models leveraging contrastive learning, using negative sampling techniques for designing more aligned context-aware retrival models, which leveraged an performance boost of over 4% on average context score, across all possible client markets there by reducing the query resolution time by 5% on average. − Worked on different fine-tuning techniques like PEFT,SFT, contrastive learning etc open source LLMs like GIST,Mistral,LLaMa for developing more context aware models for the buisness.

− Developed Sentence classifiers for multiple languages (Multilingual) achieving an average ROC-AUC of 80% for classifying sentences into categories like education, experience, etc., enhancing the search functionality significantly. − Developed a Named-Entity-Recognition System (NER) for extracting potential skills from a given job-description, this system leverages the BERT architecture for extracting important skills leveraging the atttention mechanism, and achieved an precision rate of 85% in extracting the skills. − Developed a Job-Parsing service using LLMs for scraping out the important sentences like Experience, Education etc., from the job-description, this LLM was fine-tuned using SFT and techniques like LORA and Q-LORA, was able to acheive an F1-Score of 89% . − Developed a Search LTR system for re-ranking the keyword search results, based on the frequent user-interaction, developed a reward based Multi-armed-bandit framework which would re-rank the search results based on rewards assigned. − Developed a Recommendation system ”Apply/Cart-based-recommendations” from scratch which recommends jobs based on user’s previous applications/ jobs saved in the cart, this system leverages BERT for contextual understanding of user’s and then fine tune’s a Neural Network for providing recommendations, this system achieved a CTR of 4.5% from the previous existing system which was based on string matching. − Developed a Recommendation system ”people-also-viewed” using Deep Colloborative Filtering which understands the user-user and user-item patterns and recommends jobs based on user preferences, this is a top-notch widget which achived a CTR of 5% CTR from the previous existing system which was based on Elastic Search. − Developed a Sequential Recommendation system ”browsing-based-recommendations” BERT4Rec which gives recommendations based on user’s browsing history, this system was developed using concepts of BERT which leverages the BERT for understanding the user pattern in a session and giving recommendations based on the interactions, this system achieved an CTR of 3.5 from the existing system. − Engineered Recommendation systems for making them available for all clients on large scale using Kubernetes, Kakfa and AWS Lambda, monitored the metrics by creating dashboards on Grafana using grafana query language. − Developed a Recommendation system which recommends jobs based on the blogs which the companies use, this system was developed using Open-AI and semantic search − Integrated and Fine-Tuned a Learning to Rank (LTR) model with all the Recommendation systems developed, which enhanced user engagement by providing more relevant recommendations at the top, this system improved the click rate by 7.5% than the normal recommendations which we’re not ordered based on ranking.

Oracle Generative AI Professional Certification

Oracle

Google Data Analytics Professional Certifcation

Google

Google IT Automation Professional Certifcation

Google