Shastri Vaishampayan Mohapatra

• Architected CFG-driven synthetic data generation pipelines to scale training corpora for transformer-based NLU models, improving intent classification and NER generalisation across multilingual (Hindi, English), automotive voice assistant interactions.
• Led data-centric model optimisation cycles through fine-grained error analysis and metric-driven evaluation (F1, accuracy), resulting in measurable improvements in intent disambiguation, and entity extraction performance in production systems.
• Designed and deployed annotation quality assurance frameworks using regex and rule-based validation, reducing schema violations, label noise, and strengthening training data reliability for large-scale NLP pipelines.
• Conducted linguistically informed morphological analysis for Hindi, enabling the development of robust normalisation and subword tokenisation strategies aligned with transformer tokenisation schemes.
• Engineered tokenisation and normalisation grammars to improve handling of noisy, code-mixed, and morphologically rich inputs, leading to enhanced NLU robustness and downstream interpretability.
• Contributed to improving end-to-end speech-to-intent understanding by optimising text normalisation layers bridging the ASR output and downstream NLU components, resulting in better alignment between spoken input variability and structured intent parsing.
• Collaborated across ASR, NLU, and product systems to optimise the speech-to-intent pipeline, ensuring alignment with real-world conversational distributions, long-tail queries, and edge-case behaviours.

• Collaborated with the Linguistic Project Manager and NLP researchers on various projects, including:
• Reviewing and annotating linguistic datasets to ensure data integrity for machine learning models.
• Summarising outputs from Large Language Models (LLMs) for enhanced usability.
• Providing linguistic analysis in the semantic, syntactic, and morphological domains.
• Consulting on the development of linguistic databases to support advanced NLP applications.
• Conducted research and analysis to address language-related challenges, improving AI system performance.
• Developed and implemented linguistic models and algorithms to optimise language generation and enhance AI product quality.
• Analysed AI system performance through rigorous testing, offering actionable insights for improvement.
• Collaborated with data engineering teams to collect and annotate linguistic data, ensuring accuracy in training datasets.
• Mentored junior linguists and data scientists, fostering skill development across projects.
• Executed prompt annotation and evaluation for LLM projects using ServiceNow APIs.
• Conducted sensitive data masking by recasting PII and using text normalisation and semantic transformation to train LLMs for automated dialogue generation, in line with Ethical AI and Responsible AI principles.

• Worked on building Phonemizer (a mapping tool that maps Odia phonemes with IPA) for Odia Language
• Responsible for localization, and quality data production for NLU, TTS, and ASR for AI Voice Bot, and WhatsApp chatbot for English, Hindi, and Odia in the domain of BFSI
• Worked on Mind-map for creating the Bot response Flow for voice Bot, and WhatsApp chatbot
• Worked on Conversation Design aimed at best customer experience
• Domain: BFSI (Debt Collection, Sales & Onboarding)
• Leadership Role: Domain specific POC for Quality Analysis of the data produced by the team members, and team management.

• Translation, Editing, and Review (English Oriya), (Hindi English)
• QC and QA for Translation and Voice Over (English, Hindi, Oriya)
• Domain: Academic, Medical & Health, Industry, App, News, etc.

Development of a Rule-based Multi-lingual Parser and Machine Translation System

• 
  
• Description: Designed a multi-lingual parser and MT system for translating Hindi into Oriya, English, Tamil, German, and Japanese.
• Key Contributions: Built essential resources, including a Transfer Grammar module, Mapper, and a Multilingual Concept Dictionary for accurate MT generation. Led semantic analysis, annotation, and disambiguation tasks to develop a disambiguated Universal Semantic Representation (USR) for improved translation quality.

Synthesis of Data for Conversational AI

• Developed and curated high-quality synthetic Hindi language datasets to enhance the performance of AI-driven chatbots in the BFSI sector, ensuring accurate and contextually relevant customer interactions.

Gold Data Creation for Question Data Disambiguation

• Curated gold-standard data to refine the disambiguation process of user queries, significantly enhancing the chatbot’s ability to accurately interpret and respond to complex inquiries in Hindi.

Creation of Synthesized and Gold Question-Answer Data for Banking Chatbots

• Designed and implemented a robust question-answer (QA) dataset, combining both synthesized and gold data specifically for banking chatbots. This approach facilitated seamless customer engagement and efficient resolution of banking-related queries in the BFSI domain.

• Oriya, Hindi, Sanskrit Transcription, disambiguation, Transliteration, Quality Evaluation
• Indian English pronunciation rating.