Roshan Tiwari

About Roshan Tiwari

Data Scientist/ Applied Scientist with 3+ years of experience designing, training, and deploying large scale ML and LLM systems in production. Specializes in building agentic architectures, fine tuning LLMs (LoRA, DPO, PPO), and developing enterprise grade RAG pipelines. Demonstrated ownership of end to end ML workflows — from Bayesian modeling and uncertainty aware classifiers to MLOps pipelines with automated retraining triggered by data drift — delivering multi million dollar impact through system reliability, evaluation rigor, and scalable model deployments.

Experience

1. Data Scientist

   CBRE

   Jan 2023 — Present · Dallas, TX, US

   Team Unity at CBRE is responsible for extracting information from unstructured pdf documents like leases, property tax invoices, etc. Named Entity Recognition Deployed an ensemble architecture of BERT models to extract 120 entities from pdf documents(commercial leases and amendments). Reached a weighted f1 of 0.8 with limited annotation examples by splitting low frequencyattributes between multiple models and stitching them back together using a transformer ensemble. This was a high impact projectfor CBRE’s top clients like IBM and saved the company over 50 million dollars in yearly data entry and data extraction costs. Text Summarization using LLM: Finetuned Large Language Models (Llama, Phi) using the technique quantized LORA (low rankadaptation of LLMs) to create summarizations of verbose entities like clauses, terms and conditions etc. Gen AI summarizationimproved rogue scores from 0.3 to 0.5 over prevailing techniques like seq-to-seq modelling. Retrieval Augmented Generation: Built a RAG architecture to leverage LLMs for querying large documents exceeding LLM’s contextwindow. Combined with prompting techniques, like few-shot learning and chain of thought, it enabled GPT-4o to make connectionsbetween distinct contexts and respond to queries that required critical thinking. ML Ops: Utilized Kubernetes offerings to upgrade model development lifecycle – 1. Automated and distributed hyperparametertuning on Kubernetes clusters using Katib. Automated model retraining upon drift detection via Kubeflow pipelines. Providedasync capabilities by decoupling data storage and model processing through a Kafka layer.

Education

• Vellore Institute of Technology

  Bachelor's degree, Industrial and Production Engineering

  2014 — 2018

• The University of Texas at Dallas

  Master of Science - MS, Information Technology Project Management

  2021 — 2022