Apr 2024 – Present · 2 yrs 2 mos

• Led and mentored a team of 6–8 data scientists and ML engineers, owning sprint planning, technical direction, and delivery for multiple Generative AI initiatives.
• Defined GenAI solution architecture and delivery roadmap, aligning agentic system design with business KPIs, security constraints, and client expectations.
• Spearheaded a commercialized SAP Generative AI product, automating Technical Specification generation using GPT-4o, multimodal document understanding, and custom RAG pipelines, resulting in 30–40% productivity improvement.
• Designed and delivered multi-agent GenAI systems using LangChain, LangGraph, AutoGen, and Strands, deployed on AWS and Azure, enabling deterministic orchestration via tool-based agents.
• Acted as technical owner and escalation point for LLM architecture, RAG quality, agent evaluation, and prompt governance, reducing rework and iteration cycles.
• Guided team members on LLM fine-tuning (PEFT/LoRA), quantization (4-bit), and RoPE scaling, achieving lower inference costs and improved latency in production.
• Led development and deployment of LLM inference endpoints on AWS EC2, integrating Amazon Bedrock, Azure OpenAI (GPT-4o), and IBM watsonx.ai.
• Collaborated with product owners, SMEs, and architects to translate ambiguous requirements into production-ready AI system designs.
• Conducted architecture and code reviews, mentoring team members on agentic design patterns, MLOps readiness, and enterprise AI governance.
• Defined evaluation metrics and acceptance criteria for GenAI outputs, balancing accuracy, determinism, and usability for enterprise adoption.

Jul 2022 – Apr 2024 · 1 yr 9 mos

• Generative AI Initiative:
• Fine-tune Large Language Models such as StarCoder, Llama2, and Code Llama, across diverse use cases using PEFT. Notably, the instruction-tuned StarCoder model outperformed the base model on the Huggingface Leaderboard.
• Spearheaded the development of a Langchain-based application, harnessing VectorDB's capabilities and a fine tuned StarCoder model to seamlessly translate natural language into SQL queries and visualize results through a Streamlit user interface.
• Implemented prompt engineering and fine-tuned few-shot techniques to enhance the performance of Large Language Models (LLMs) effectively.
• Skillfully fine-tuned hyperparameters for LLMs, tailoring them to the specific requirements of each use case to deliver optimal results.
• Led the development of an inference endpoint for various Large Language Models (LLMs) and achieved a successful deployment on an AWS EC2 instance.
• Demonstrated a dynamic approach by integrating Azure GPT-3 and IBM WatsonX inferencing endpoints into Langchain applications, enhancing their capabilities.
• Implemented innovative RoPE scaling techniques to expand the context length of Llama2 and Code Llama for more effective inferencing.
• Successfully deployed 4-bit quantized versions of the models for inferencing to reduce operating costs.
• CollabAI, IBM Research:
• Developed and trained ARX, ARIMAX & Prophet models for forecasting time series sales inventory data.
• Trained and hypertuned AR, Naive, LSTM, RF, XGBoost models for the inventory forecasts.
• Developed an ensemble machine learning model for sales inventory forecasts with a MAPE of 0.19.
• Leveraged CPLEX, qpsolver & Pulp to programmatically solve linear & quadratic optimization problems

Jun 2019 – Jul 2022 · 3 yrs 1 mo

• Acoustic Analysis:
• Developed gmm, dcase & YoloV5 machine learning models for welding audio files & trained a classifier.
• Developed and trained a GoogLeNet classifier in IBM Maximo Visual Inspection suite.
• The classifier had an accuracy of ~90%, and was able to process an audio in a time 10% of file length.
• Designed and developed a Python Flask app to classify an audio file against a ML classifier of choice.
• Sandvik Optimine Analytics:
• Developed script for data validation of IoT streaming data to generate flags for mining equipment.
• Used IBM Watson Studio to deploy, monitor and improve upon various data science jobs.
• Published an asset to IBM Lighthouse Repository for automating replication of database objects.
• Pearson UK:
• Developed an application to extract and transform data from various SQL & NoSQL databases.
• Developed a Python application to securely PGP encrypted data over a SFTP server.
• Created Docker images of various applications and deployed them to a Kubernetes cluster.
• Leveraged various parallel computing libraries for Python such as Dask, Vaex and Modin with Ray.
• Sensor Detection AAHK:
• Analyzed the intensity of vibrations of sensors to detect the quality of pillars to estimate the degradation.
• Developed a K-means classification model on sensor data for analyzing the quality.