May 2024 – Aug 2024 · 3 mos · New York, United States · Hybrid

• Designed high-level and low-level components of an AI assistant for equity research and financial market analysis. I was the first AI hire, owned and built everything from scratch.
• Developed end-to-end components of a keyword-guided retrieval and search engine. Used BM25, Elastic Search, dense vector search, fine-tuned Splade, and ColBERT models for query augmentation and candidate reranking.
• Built an advanced agentic RAG framework with CI/CD pipelines for query understanding, document ingestion, and knowledge graph building to support Q&A over the past 10 years of financial documents from S&P 2000 companies.
• Developed and containerized microservices to fine-tune and serve LLMs (LLaMa-3, Phi-3 and Mixtral) as APIs using Docker and Kubernetes. Employed techniques like quantization, flash attention, LoRA, data and model parallelism for efficient fine-tuning on GPUs. Utilized vLLM, Paged-attention, Kafka, and Redis for faster LLM inference service.

Jan 2022 – Aug 2023 · 1 yr 7 mos

• Designed and implemented Model inference service, Regression and load-testing framework to integrate Large-Language-Models and Generative AI capabilities into Instabase platform. https://aihub.instabase.com/
• Designed and implemented an async Model serving platform, which combines async workers, rabbit-MQ, two levels of caching, and sticky routing techniques to improve inference time by 40% on a compute-limited Kubernetes environment.
• Devised a model drift detection pipeline that uses layout change detectors to detect format drift and uses Least-Squares Density Difference and Maximum Mean Discrepancies on document embedding to catch content drift in the documents.

Aug 2020 – Jan 2022 · 1 yr 5 mos · Bengaluru, Karnataka, India

• Implemented the entity embedding technique for sparse categorical values, which improved the accuracy of the existing forecasting systems by 17% and reduced the inference time by 15%. This also reduced the effort required for feature engineering.
• Tweaked the gradient boosting algorithm to incorporate the feedback of the distribution center’s manager while calculating errors to form decision trees and implemented weighted loss to penalize misclassifications of perishable items such as fruits and dairy products more heavily.

May 2019 – Aug 2019 · 3 mos

• Developed a Google Chrome extension that uses natural language processing (NLP) to detect toxic comments, clickbait, and fake news on news websites and social media platforms like Facebook and Twitter. Trained a stance detection model for fake news classification and implemented a backend service using Flask for model inference. Designed a three-level cache system to optimize latency and avoid duplicate inference requests.

May 2019 – Jul 2019 · 2 mos · Hyderabad Area, India

• Implemented passage ranking algorithm to boost the accuracy of in-house Question and Answer services’ model on some client data sets.Created a framework to train AllenNLP’s machine reading comprehension models on in-house medical data as part of a client’s POC and proposed its integration with the homegrown Chat Bot.