May 2025 – Aug 2025 · 3 mos · Seattle, Washington, United States · On-site

• Replaced Visa's HDFS-based batch reporting with a low latency, high throughput ClickHouse pipeline, achieving 58% faster processing speeds (29,000 vs 18,000 records/sec) and 37% higher throughput
• Built a scalable ETL pipeline connecting Visa's SQL server to ClickHouse to enable sub-second query performance, materialized views for pre-computed aggregations, and integration with Tableau
• Leveraged ClickHouse's columnar storage and vectorized execution engine to eliminate Hadoop's multi-hop latency and ensure consistent performance across large-scale usage datasets.
• Built a time series forecasting pipeline using temporal features and an LSTM-based classifier to predict merchant attrition 3-8 months in advance, enabling proactive intervention; the model achieved 96% precision for "Will stay" and 76% recall for "Will Churn" predictions
• Used SHAP-based interpretability analysis, revealing the LSTM model's over-reliance on payment volume features, leading to a redesign focused on behavioral signals and multi-dimensional churn indicators
• Developed an enhanced two-step XGBoost pipeline that improved churn recall to 90% and enabled month-specific churn forecasting (eg., predicting whether a merchant would churn in 3,4,5, or 6 months).

Jan 2025 – Nov 2025 · 10 mos · New York, New York, United States · Remote

• With mentors from PwC alongside mentors from Columbia University, Built a modular benchmarking framework to measure inference latency, memory, energy and carbon emissions across sweeps of model type, quantization(fp16/int8/int4), batch size, and I/O length
• Integrated Zeus GPU power monitoring with automated experiment logging and W&B visualization
• Benchmarked LLaMA-2, Deepseek-Distill, SparseBERT, and Switch Transformer on Aplaca and GSM8K, capturing performance trends across dense, sparse and MOE architectures
• Identified batch size as the dominant factor: increasing from 1->4 cut latency by 60% and reduced energy/token by 50%
• Sparse and MoE models achieved lowest energy/token across both GPUs
• A100 delivered 10x lower latency and 75% less total energy than L4 for identical workloads

Dec 2024 – Feb 2026 · 1 yr 2 mos · New York, New York, United States · On-site

• Built a multi-agent AI driven simulation of Cambridge KET and PET English speaking exams using the Agent-First-Organization framework, orchestrating interactions between a GPT-4o based AI Examiner, AI Student, and a real student interface
• Engineered the exam flow across three stages (personal interview, topic-based discussion, and follow-up) by designing custom worker nodes and task graphs that controlled turn-taking, topic progression, and difficulty-based response generation
• Incorporated OpenAI's embedding model to semantically match student responses against a curated rubric of common grammar and vocabulary errors for accurate, rubric-aligned scoring
• Extended the PET system to support image-based prompts by designing visual-based interaction modules and embedding captioned diagrams, mimicking real exam visuals.
• Developing an automated patent description generation system leveraging Claude and OpenAI APIs with integrated style transfer capabilities
• Building end-to-end patent figure analysis pipeline leveraging multimodal AI to automatically identify identify components validate technical accuracy, and generate formal patent descriptions from technical drawings using Chain-of-Thought prompting strategies

May 2023 – Aug 2023 · 3 mos · San Jose, California, United States · On-site

• Did training on Internet Key Exchange(IKE), IPsec, SD-WAN, particularly with reference to Silver Peak's EdgeConnect and Orchestrator
• Added support for the EdgeConnect to be in responder-only mode during IKE processing for IPsec tunnels in SD-WAN including Command Line Interface and backend development in C
• Ran multiple trials with different Strongswan configurations to evaluate potential solutions
• Worked on improvements to EdgeConnect IPsec Tunnel Debuggability from the Orchestrator by adding IKE and IPsec tunnel logs with filter patterns in all stages of standard IPsec and IKEless tunnel bring up, including key exchange and datapath in C
• Added anti-replay state information for tunnel troubleshoot in C language

Sep 2022 – Aug 2024 · 1 yr 11 mos · On-site

• Applied YOLOv5 and GroundingDINO for crop disease detection, improving localization accuracy by 30% on benchmark datasets, and designed a preprocessing pipeline with Gaussian filtering and resolution normalization to enhance model input quality
• Optimized training pipeline through data augmentation and learning rate scheduling, reducing false positives by 25% across varied crop disease images

May 2022 – Oct 2022 · 5 mos · Massachusetts, United States

• Worked on distributed velsyn which is a multi-process implementation of the Velsyn compiler
• Designed and developed Dominator Propagation algorithm which is used for logic optimization
• Created testcases using Verilog to test full implementation of the code
• Programmed in C using Linux and debugged the design using Gdb in distributed systems
• Redesigned a connectivity comparison Python script used to compare netlist connectivity between - sequential and distributed compile of designs to ensure correct connectivity across cores
• Connectivity script takes less than 20 minutes on a design that originally did not finish for hours