cv ML SWE CV

Master of Science in Computer Science

• Concentration: Artificial Intelligence
• Worked as a Research Assistant at CILVR lab, advised by Shauli Ravfogel, Jackson Petty, and Tal Linzen.

Bachelor of Technology in Computer Science & Engineering

• Specialization: Machine Intelligence and Data Science
• Received the Undergraduate Researcher Award for my work in the field of Machine Learning.
• Worked as a Research Assistant at the Center for Cloud Computing & Big Data, advised by KV Subramaniam.

Applied Research Scientist Intern

Advisor: Srijith Rajamohan

• Introduced Precision–Cache Hit Ratio (P-CHR) and Calibration Retention Rate (CRR), two cache-aware metrics for Redis LangCache, reframing semantic-cache model selection as a calibration problem rather than a ranking one.
• Curated and open-sourced LangCache SentencePairs (v1-v3), a large-scale dataset family with 1M to 40M sentence pairs across diverse paraphrase, STS, QA, and adversarial sources for robust semantic-caching fine-tuning.
• Fine-tuned and deployed LangCache-Embed-v3, a domain-specific bi-encoder trained with an ArcFace contrastive objective, leading all open-source retrievers offline and outperforming larger general-purpose models even without re-ranking.
• Open-sourced the LangCache ReRanker v1/v2 cross-encoder families with BCE and MNRL objectives at two training scales, isolating the effect of objective vs. scale on calibration and enabling application-specific behavior.
• Built a comprehensive evaluation framework for customer onboarding with a large-scale study of customer data and model baselines across offline and deployment settings, using a two-stage RedisVL K-NN retrieval and re-ranking pipeline.
• Demonstrated that the highest-PR-AUC models are often the worst in deployment, and that re-ranking rarely improves a strong domain retriever, overturning the standard practice in semantic-cache model selection.
• Supported downstream integration and development of LMCache by building prototypes and benchmarking performance with Redis as an in-memory KV store, demonstrating latency and throughput gains.

Research Assistant

Advisors: Shauli Ravfogel, Tal Linzen

• Investigated arithmetic circuit dynamics in language models when operators are redefined in-context by analyzing activation representations and attention patterns across transformer layers.
• Trained tiny (~3M params) LLaMA-style transformers on arithmetic operations under operator overloading to test whether transformers reuse circuits across semantically related tasks.
• Developed an activation-extraction and causal-attribution pipeline pairing neuron selection with Activation Patching, Direct Logit Attribution, and Fourier probes to separate causally necessary components from merely predictive ones.
• Identified and characterized a two-tier circuit structure across a model’s depth and breadth, showing transformers organize computation around semantic operations rather than surface symbols.

Research Assistant

Advisors: Jackson Petty, Tal Linzen

• Evaluated LLMs on compositional generalization and instruction synthesis by studying their ability to translate synthetic Context-Free Grammars (CFGs) into conforming strings.
• Analyzed model outputs in few- and zero-shot settings to assess grammatical conformity and uncover generation strategies used during translation.

Big Data and Applied AI Engineer

• Instruction fine-tuned LLMs like Mistral and Llama-2 on-prem to enable secure and cost-effective AI solutions such as translation and RAG for engineers and customers, cutting third-party dependency costs by 30%.
• Led the initiative to build a novel pre-training algorithm for conversational data using PyTorch and HuggingFace, achieving a 40% performance gain over standard approaches at benchmark fine-tuning tasks.
• Developed the Webex Contextual Search engine and improved searching, ranking, recommendations, and topic modeling by 75% with <10% increased overhead latency.
• Integrated OpenAI APIs and on-prem LLMs with the Webex AI Assistant for 15M+ worldwide users to add auto-replies, summarization, querying, and action-item extraction to message threads and meeting transcripts.
• Developed and deployed streaming jobs in Scala and Flink to process 1M+ reports/min and compute 1200+ real-time metrics from Calls and Meetings.
• Applied statistical modeling techniques to investigate and report media quality insights to downstream consumers, reducing errors by 30% and analysis time by 15 hrs/week per team member.
• Led the development of real-time (<1 min) auditing pipelines using Kafka and Python to ensure per-minute data consistency between streaming jobs and Iceberg and Pinot data stores, reducing manual effort by >80%.
• Built graphs and dashboards on the Webex Media Quality Analytics Dashboard using Grafana and Kibana to set up alerts and KPIs for 20,000+ clients and customers.

Big Data Engineering Intern

• Migrated the Meetings Analytics Engine from Java and Spark to Scala and Flink to scale up to 1M+ reports/min and significantly improve real-time report generation by over 40%.
• Built VideoMesh Developer APIs using Java and globally rolled them out for 30,000+ enterprises with customer-facing applications.

Applied Research Scientist Intern

Advisors: Anay Majee, Anbumani Subramanian

• Explored Few-Shot Learning Object Detection (FSOD) techniques to reduce catastrophic forgetting in constrained and heterogeneous driving environments.
• Investigated and designed novel representation learning and attention mechanisms to learn inter/intra-object relationships using PyTorch.
• Outperformed existing approaches at the time on base and novel classes by 0.2 mAP and 3 mAP on the Few-Shot India Driving Dataset, a benchmark for FSOD.

Research Assistant

Advisor: KV Subramaniam

• Compiled and used TailBench to simulate and profile application loads, monitor performance, and analyze results.
• Explored ways to reduce tail latencies in latency-critical applications such as translation and image recognition.

Second Place out of 20+ teams at Webex Analytics Datathon 2024

Cisco Systems

Containerized and deployed a self-sufficient, on-prem and quantized LLM-RAG pipeline to assist engineers with engineering queries and incident resolution.

Ranked #1 Internationally out of 300+ teams at the Webex IDEA Hackathon 2023

Cisco Systems

Integrated OpenAI LLM APIs with the Webex Assistant to enable summarization of message threads, media and transcripts (demo). Developed thread-related user actions like searching, grouping and sorting across Webex. Assisted in globally rolling out these features worldwide.

Ranked #1 regionally and Top 20 Internationally out of 300+ teams at the Webex Playtime Hackathon 2023

Cisco Systems

Developed the Webex Contextual Search engine using novel conversational representation learning techniques and displayed significant improvement in searching, ranking and recommendations.

Undergraduate Researcher Award

PES University

Awarded for work in the field of Machine Learning.

3x Scholarship Recipient (Prof. CNR Rao, MRD & DAC Scholarship Awards)

PES University

For being in the top 20% among 900+ students.

Finalist at Intel Technovation, Flipkart, IBM, and IISc Hackathons

PES University

Placed among the top 200+ teams.

National newspaper coverage for proposing the model to track garbage collection in Bengaluru

CMR National Public School

Received extensive coverage and recognition for developing an Android app to track and schedule garbage collection in Bengaluru. The currently implemented model was based on our designs and proposals made to the BBMP. Coverage: The Hindu, India Today, Times of India.

Teaching Assistant, CS322: Big Data

PES University

Faculty: KV Subramaniam, Prafullata K Auradkar, Animesh Giri. Designed and graded coursework, assignments and projects, and delivered hands-on sessions on Hadoop and Spark for a class of 600+ enrolled students for the undergraduate Big Data course.

Research Advisor, Summer Internship — Center for Cloud Computing & Big Data

PES University

Led the mechanistic interpretability group, advising 2 teams on distinct projects that train small transformers to study grokking, activation patching, and circuit-level learning and inference dynamics.

Speaker, Guest Lecture — Building Foundation Models using Transformers

PES University

Delivered a guest lecture to undergraduate students on the advancements in representation learning techniques for language and highlighted the importance of interdisciplinary research.