Profile Picture

Aman Tiwari

Applied Research Scientist

Email +91-8755359531 LinkedIn GitHub Google Scholar

Work Experience

ServiceNow

Applied Research Scientist | May 2024 – Present

At ServiceNow, I work on various NLP projects for multiple clients, focusing on data curation, training SLMs and LLMs, and evaluating their performance. I have developed in-house frameworks to enhance core functionalities.

Text-to-Anything Project

  • Developed a high-precision synthetic data generation pipeline supporting multiple Text-to-X tasks, such as Text-to-Code, Text-to-Cypher (Neo4j), and Text-to-SQL.
  • Trained various models and adaptors for LLMs to enhance performance across tasks.
  • Achieved a notable improvement in task accuracy, with Text-to-Cypher accuracy rising from 20-30% to 65-70%, outperforming open-source models.

Change Risk

  • Created an ensemble model framework that integrates text and tabular data to evaluate risks in change management.
  • Incorporated SHAP values and fine-tuned LLMs to enhance model explainability and stakeholder transparency.

Embeddings Adapter Project

  • Designed a framework for synthetic data generation to produce search queries for various corpora, boosting model performance.
  • Trained adaptors on top of embedding models like BERT and E5 for improved search retrieval and classification.

Thoughtworks

Senior Data Scientist | September 2020 – May 2024

At Thoughtworks, I led the development of AI and NLP solutions, focusing on domain-specific models and revenue forecasting, which significantly enhanced client outcomes.

Aalap - Indian Legal LLM

  • Directed the creation of a domain-specific Indian legal LLM by assembling an SFT dataset and fine-tuning an 8-billion-parameter Mistral model for specialized legal tasks.

Jugalbandi Project - Conversational AI

  • Developed a Post-RAG architecture for generative AI solutions capable of speech interaction.
  • Implemented advanced Gen-AI chatbots, including Jugalbandi Bot, which is in the POC stage for potential deployment by NGOs and government bodies.

Revenue Forecasting Initiative

  • Designed proprietary forecasting models tailored for regional revenue predictions, achieving a 25% improvement in accuracy over existing models.

Indian Legal NLP Research

  • Developed "opennyai," an open-source library for integrating proprietary legal NLP models.
  • Led the creation of initial NLP models for the Indian legal domain, including NER, rhetorical roles, and summarization.
  • Formulated and executed strategies for data collection to support legal NLP model development.

Open-source Speech Models

  • Built advanced speech recognition and TTS models for 15+ Indian languages, achieving a WER of less than 10% and high MOS scores.
  • Designed data collection methods, including speaker clustering and language identification systems.

Skills

Programming

Python

Machine Learning

Supervised & Unsupervised Learning Deep Learning Transfer Learning Ensemble Methods Gradient Boosting (XGBoost, LightGBM, CatBoost) Neural Networks

Deep Learning

CNNs RNNs LSTMs Transformer Models (BERT, GPT) Attention Mechanisms

Natural Language Processing

LLM’s finetuning and alignment Text Classification NER Machine Translation Sequence-to-Sequence Models Word Embeddings Speech Recognition

Data Science

Data Cleaning Feature Engineering Data Visualization Predictive Modelling Statistical Modelling Time Series Analysis A/B Testing

Frameworks & Libraries

TensorFlow PyTorch Keras Scikit-learn Pandas NumPy SciPy Hugging Face OpenCV NLTK SpaCy

MLOps & Deployment

Docker Kafka CI/CD ONNX Fast API MLflow Model Monitoring

Cloud Platforms

AWS GCP Azure

Education

Master of Science – Data Science

University of Arizona

2022 - 2024

Bachelor of Technology – Computer Science

DIT University

2016 - 2020

Publications / Papers

Note: The above list includes some of my key publications. For a complete list, please visit my Google Scholar profile.

Important Open-Source Projects

Opennyai

This project creates a streamlined NLP pipeline for Indian legal documents, featuring Named Entity Recognition (NER), rhetorical role identification, and summarization. The pipeline is carefully designed and trained, with rigorous data curation and annotation, and evaluated using robust metrics.

Jugalbandi

Jugalbandi is a free and open platform that combines the power of Large Language Models such as ChatGPT and Indian language translation models such as those under the Government of India's Bhashini mission to power conversational AI solutions in any domain.

Aalap: A 32K context length Indian legal LLM

Aalap (Assistant for Legal and Paralegal functions in India) is an instructions fine-tuned version of Mistral 7B that can perform specific legal tasks in the Indian context. This research model intends to show that we can develop tasks for the legal domain and teach LLMs to do them at an affordable cost.